Lecture Notes in Mathematics Editors: J.-M. Morel, Cachan B. Teissier, Paris
For further volumes: http://www.springer.com/series/304
2023
Kreˇsimir Veseli´c
Damped Oscillations of Linear Systems A Mathematical Introduction
123
Kreˇsimir Veseli´c Fernuniversit¨at Hagen Fakult¨at f¨ur Mathematik und Informatik Universit¨atsstr. 11 58084 Hagen Germany
[email protected] ISBN 978-3-642-21334-2 e-ISBN 978-3-642-21335-9 DOI 10.1007/978-3-642-21335-9 Springer Heidelberg Dordrecht London New York Lecture Notes in Mathematics ISSN print edition: 0075-8434 ISSN electronic edition: 1617-9692 Library of Congress Control Number: 2011933133 Mathematics Subject Classification (2011): 15-XX, 47-XX, 70-XX c Springer-Verlag Berlin Heidelberg 2011 This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilm or in any other way, and storage in data banks. Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in its current version, and permission for use must always be obtained from Springer. Violations are liable to prosecution under the German Copyright Law. The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. Cover design: deblik, Berlin Printed on acid-free paper Springer is part of Springer Science+Business Media (www.springer.com)
To my wife.
Foreword
The theory of linear damped oscillations has been studied for more than 100 years and is still of vital interest to researchers in Control Theory, Optimization, and computational aspects. This theory plays a central role in studying the stability of mechanical structures, but it has applications to other fields such as electrical network systems or quantum mechanical systems. We have purposely restricted ourselves to the basic model leaving aside gyroscopic effects and free rigid body motion. In contrast, the case of a singular mass matrix is analysed in some detail. We spend quite a good deal of time discussing underlying spectral theory, not forgetting to stress its limitations as a tool for our ultimate objective – the time behaviour of a damped system. We have restricted ourselves to finite dimension although we have attempted, whenever possible, to use methods which allow immediate generalisation to the infinite-dimensional case. Our text is intended to be an introduction to this topic and so we have tried to make the exposition as self-contained as possible. This is also the reason why we have restricted ourselves to finite dimensional models. This lowering of the technical barrier should enable the students to concentrate on central features of the phenomenon of damped oscillation. The introductory chapter includes some paragraphs on the mechanical model in order to accommodate readers with weak or no background in physics. The text presents certain aspects of matrix theory which are contained in monographs and advanced textbooks but may not be familiar to the typical graduate student. One of them is spectral theory in indefinite product spaces. This topic receives substantial attention because it is a theoretical fundament for our model. We do not address numerical methods especially designed for this model. Instead we limit ourselves to mention what can be done with most common matrix algorithms and to systematically consider the sensitivity, that is, condition number estimates and perturbation properties of our topics. In our opinion numerical methods, in particular invariant-subspace reduction vii
viii
Foreword
for J-symmetric matrices are still lacking and we have tried to make a case for a more intensive research in this direction. Our intention is to take readers on a fairly sweeping journey from the basics to several research frontiers. This has dictated a rather limited choice of material, once we ‘take off’ from the basics. The choice was, of course, closely connected with our own research interests. In some cases we present original research (cf. e.g. Chap. 19), whereas we sometimes merely describe an open problem worth investigating – and leave it unsolved. This text contains several contributions of our own which may be new. As a rule, we did not give any particular credit for earlier findings due to other authors or ourselves except for more recent ones or when referring to further reading. Our bibliography is far from exhaustive, but it contains several works offering more detailed bibliographical coverage. The text we present to the reader stems from a synonymous course I have taught for graduate and post-doc students of the Department of Mathematics, University of Osijek, Croatia in the summer term 2007/2008. It took place at the Department of Mathematics at the University of Osijek and was sponsored by the program ‘Brain gain visitor’ from the National Foundation for Science, Higher Education and Technological development of the Republic of Croatia. Due to its origin, it is primarily designed for students of Mathematics but it will be of use also to engineers with enough mathematical background. Even though the text does not cover all aspects of the linear theory of damped oscillations, I hope that it will also be of some help to the researchers in this field. In spite of a good deal of editing the text still contains some remnants of its oral source. This pertains to the sometimes casual style more suited to a lecture than to a monograph – as was the original aim of this work. Bibliographical Notes and Remarks are intended to broaden the scope by mentioning some other important directions of research present and past. According to our bibliographical policy, when presenting a topic we usually cite at most one or two related works and refer to their bibliographies. It is a pleasure to thank the participants in the course, in particular professor N. Truhar and his collaborators for their discussions and support during the lectures in Osijek. The author was very fortunate to have obtained large lists of comments and corrections from a number of individuals who have read this text. These are (in alphabetic order): K. Burazin (Osijek), L. Grubiˇsi´c (Zagreb), D. Kressner (Z¨ urich), P. Lancaster, (Calgary), H. Langer (Vienna), I. Mati´c (Osijek), V. Mehrmann (Berlin), M. Miloloˇza Pandur (Osijek), B. Parlett (Berkeley), I. Naki´c (Zagreb), A. Suhadolc (Ljubljana), I. Veseli´c (Chemnitz), A. Wiegner (Hagen) and, by no means the least, the anonymous referees. Many of them went into great detail and/or depth. This resulted in substantial revisions of the whole text as well as in significant enlargements. To all of them I am deeply obliged. Without their generous support and encouragement this text would hardly deserve to be presented to the public. The financial help of both National Foundation for
Foreword
ix
Science, Higher Education and Technological development of the Republic of Croatia and the Department of Mathematics, University of Osijek, not forgetting the warm hospitality of the latter, is also gratefully acknowledged. Osijek/Hagen July 2008–December 2010
K. Veseli´c
Contents
1
The 1.1 1.2 1.3 1.4
Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Newton’s Law .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Work and Energy .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . The Formalism of Lagrange .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Oscillating Electrical Circuits. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1 3 4 6 12
2
Simultaneous Diagonalisation (Modal Damping) . . . . . . . . . . . 2.1 Undamped Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Frequencies as Singular Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3 Modally Damped Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15 15 19 20
3
Phase Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.1 General Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2 Energy Phase Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
23 23 24
4
The Singular Mass Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
29
5
‘Indefinite Metric’ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.1 Sylvester Inertia Theorem.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
39 41
6
Matrices and Indefinite Scalar Products . . . . . . . . . . . . . . . . . . . . .
49
7
Oblique Projections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
55
8
J-Orthogonal Projections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
61
9
Spectral Properties and Reduction of J-Hermitian Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
67
xi
xii
Contents
10 Definite Spectra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
73
11 General Hermitian Matrix Pairs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
89
12 Spectral Decomposition of a General J-Hermitian Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93 12.1 Condition Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105 13 The Matrix Exponential . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113 14 The Quadratic Eigenvalue Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 121 15 Simple Eigenvalue Inclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129 16 Spectral Shift . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135 17 Resonances and Resolvents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139 18 Well-Posedness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143 19 Modal Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145 20 Modal Approximation and Overdampedness . . . . . . . . . . . . . . . . 159 20.1 Monotonicity-Based Bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162 21 Passive Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21.1 More on Lyapunov Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21.2 Global Minimum of the Trace . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21.3 Lyapunov Trace vs. Spectral Abscissa . . . . . . . . . . . . . . . . . . . . . . . .
167 170 172 178
22 Perturbing Matrix Exponential . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185 23 Notes and Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203 Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
Introduction
Here is some advice to make this book easier to read. In order to check/deepen the understanding of the material and to facilitate independent work the reader is supplied with some thoroughly worked-out examples as well as a number of exercises. Not all exercises have the same status. Some of them are ‘obligatory’ because they are quoted later. Such exercises are either easy and straightforward or accompanied by hints or sketches of solution. Some just continue the line of worked examples. On the other end of the scale there are some which introduce the reader to research along the lines of the development. These are marked by the word ‘try’ in their statement. It is our firm conviction that a student can fully digest these Notes only if he/she has solved a good quantity of exercises. Besides examples and exercises there are a few theorems and corollaries the proof of which is left to the reader. The difference between a corollary without proof and an exercise without solution lies mainly in their significance in the later text. We have tried to minimise the interdependencies of various parts of the text. What can be skipped on first reading? Most of the exercises. Almost none of the examples. Typically the material towards the end of a chapter. In particular • Chapter 10: Theorem 10.12 • Chapter 12: Theorem 12.17 • Any of Chaps. 19–22 Prerequisites and terminology. We require standard facts of matrix theory over the real or complex field Ξ including the following (cf. [31, 58]). • Linear (in)dependence, dimension, orthogonality. Direct and orthogonal sums of subspaces. • Linear systems, rank, matrices as linear maps. We will use the terms injective/surjective for a matrix with linearly independent columns/rows. xiii
xiv
Introduction
• Standard matrix norms, continuity and elementary analysis of matrixvalued functions. • Standard matrix decompositions such as – Gaussian elimination and the LU-decomposition A = LU , L lower triangular with unit diagonal and U upper triangular. – Idem with pivoting A = LU Π, L, U as above and Π a permutation matrix, corresponding to the standard row pivoting. – Cholesky decomposition of a positive definite Hermitian matrix A = LL∗ , L lower triangular. – QR-decomposition of an arbitrary matrix A = QR, Q unitary and R upper triangular. – Singular value decomposition of an arbitrary matrix A = U ΣV ∗ , U, V unitary and Σ ≥ 0 diagonal. – Eigenvalue decomposition of a Hermitian matrix A = U ΛU ∗ , U unitary, Λ diagonal. – Simultaneous diagonalisation of a Hermitian matrix pair A, B (the latter positive definite) Φ∗ AΦ = Λ,
Φ∗ BΦ = I,
Λ diagonal.
– Schur (or triangular) decomposition of an arbitrary square A = U T U ∗ , U unitary, T upper triangular. – Polar decomposition of an arbitrary m × n-matrix with m ≥ n √ A = U A∗ A with U ∗ U = In . • Characteristic polynomial, eigenvalues and eigenvectors of a general matrix; their geometric and algebraic multiplicity. • (Desirable but not necessary) Jordan canonical form of a general matrix. Further general prerequisites are standard analysis in Ξ n as well as elements of the theory of analytic functions. Notations. Some notations were implicitly introduced in the lines above. The following notational rules are not absolute, that is, exceptions will be made, if the context requires them. • Scalar: Lower case Greek α, λ, . . .. • Column vector: Lower case Latin x, y, . . ., their components: xj , yj , . . . (or (x)j , (y)j , . . .) • Canonical basis vectors: ej as (ej )k = δjk . • Matrix order: m × n, that is, the matrix has m rows and n columns, if a matrix is square then we also say it to be of order n. • The set of all m × n-matrices over the field Ξ: Ξ m,n
Introduction
xv
• Matrix: Capital Latin A, B, X, . . ., sometimes capital Greek Φ, Ψ, . . .; diagonal matrices: Greek α, Λ, . . . (bold face, if lower case). By default the order of a general square matrix will be n, for phase-space matrices governing damped systems this order will mostly be 2n. • Identity matrix, zero matrix: In , 0m,n, 0n ; the subscripts will be omitted whenever clear from context. • Matrix element: Corresponding lower case of the matrix symbol A = (aij ). • Block matrix: A = (Aij ). • Matrix element in complicated expressions: ABC = ((ABC)ij ). • Diagonal matrix: diag(a1 , . . . , an );block Diagonal matrix: diag(A1 , . . . , Ap ). • Matrix transpose: AT = (aji ). • Matrix adjoint: A∗ = (aji ). • Matrix inverse and transpose/adjoint: (AT )−1 = A−T , (A∗ )−1 = A−∗ . • The null space and the range of a matrix as a linear map: N (A), R(A). • Spectrum: σ(A). • Spectral radius: spr(A) = max |σ(A)|. • Spectral norm: A = spr(A∗ A) and condition number κ(A) = −1 AA . • Euclidian norm: AE = Tr(A∗ A) and condition number κE (A) = −1 AE A E . In the literature this norm is sometimes called the Frobenius norm or the Hilbert-Schmidt norm. • Although a bit confusing the term eigenvalues (in plural) will mean any sequence of the complex zeros of the characteristic polynomial, counted with their multiplicity. Eigenvalues as elements of the spectrum will be called spectral points or distinct eigenvalues.
Chapter 1
The Model
Small damped oscillations in the absence of gyroscopic forces are described by the vector differential equation 𝑀𝑥 ¨ + 𝐶 𝑥˙ + 𝐾𝑥 = 𝑓 (𝑡).
(1.1)
Here 𝑥 = 𝑥(𝑡) is an ℝ𝑛 -valued function of time 𝑡 ∈ ℝ; 𝑀, 𝐶, 𝐾 are real symmetric matrices of order 𝑛. Typically 𝑀, 𝐾 are positive definite whereas 𝐶 is positive semidefinite. 𝑓 (𝑡) is a given vector function. The physical meaning of these objects is 𝑥𝑗 (𝑡) position or displacement 𝑀
mass
𝐶
damping
𝐾
stiffness
𝑓 (𝑡)
external force
while the dots mean time derivatives. So, any triple 𝑀, 𝐶, 𝐾 will be called a damped system, whereas the solution 𝑥 = 𝑥(𝑡) is called the motion or also the response of the linear system to the external force 𝑓 (𝑡). In this chapter we will describe common physical processes, governed by these equations and give an outline of basic mechanical principles which lead to them. It is hoped that this introduction is self-contained enough to accomodate readers with no background in Physics. Example 1.1 As a model example consider the spring-mass system like the [ ]𝑇 one in Fig. 1.1. Here 𝑥 = 𝑥1 ⋅ ⋅ ⋅ 𝑥𝑛 where 𝑥𝑖 = 𝑥𝑖 (𝑡) is the horizontal displacement of the 𝑖th mass point from its equilibrium position and
K. Veseli´ c, Damped Oscillations of Linear Systems, Lecture Notes in Mathematics 2023, DOI 10.1007/978-3-642-21335-9 1, © Springer-Verlag Berlin Heidelberg 2011
1
2
1 The Model
k1
k2
1
k3
2
k4
3
c4
d4
4
Fig. 1.1 Oscillator ladder
𝑀 = diag(𝑚1 , . . . , 𝑚𝑛 ), 𝑚1 , . . . , 𝑚𝑛 > 0 ⎡ ⎤ 𝑘1 + 𝑘2 −𝑘2 ⎢ ⎥ ⎢ −𝑘2 𝑘2 + 𝑘3 . . . ⎥ ⎢ ⎥ ⎢ ⎥ .. .. .. 𝐾 =⎢ ⎥, . . . ⎢ ⎥ ⎢ ⎥ . . .. .. ⎣ −𝑘𝑛 ⎦ −𝑘𝑛 𝑘𝑛 + 𝑘𝑛+1
(1.2)
𝑘1 , . . . 𝑘𝑛+1 > 0
(1.3)
(all void positions are zeros) and 𝐶 = 𝐶𝑖𝑛 + 𝐶𝑜𝑢𝑡 , ⎡
𝐶𝑖𝑛
𝑐1 + 𝑐2
⎢ ⎢ −𝑐2 ⎢ ⎢ =⎢ ⎢ ⎢ ⎣
(1.4) ⎤
−𝑐2
⎥ . ⎥ 𝑐2 + 𝑐3 . . ⎥ ⎥ .. .. .. ⎥, . . . ⎥ ⎥ .. .. . . −𝑐𝑛 ⎦ −𝑐𝑛 𝑐𝑛 + 𝑐𝑛+1
𝐶𝑜𝑢𝑡 = diag(𝑑1 , . . . , 𝑑𝑛 ),
𝑐1 , . . . 𝑐𝑛+1 ≥ 0
𝑑1 , . . . , 𝑑𝑛 ≥ 0
(1.5)
(1.6)
In the case of Fig. 1.1 we have 𝑛 = 4 and 𝑘5 = 0,
𝑐2 = 𝑐3 = 0,
𝑑1 = 𝑑2 = 𝑑3 = 0,
𝑑4 > 0.
Various dynamical quantities such as ‘force’, ‘work’, ‘energy’ will play an important role in the course of these Notes. We shall next give a short compact overview of the physical background of (1.1).
1.1 Newton’s Law
1.1
3
Newton’s Law
We will derive (1.1) from Newton’s law for our model example. The coordinate 𝑥𝑖 of the 𝑖th mass point is measured from its static equilibrium position, that is, from the point where this mass is at rest and 𝑓 = 0. Newton’s second law of motion for the 𝑖th particle reads 𝑚𝑖 𝑥 ¨𝑖 = 𝑓𝑖𝑡𝑜𝑡 , where 𝑓𝑖𝑡𝑜𝑡 is the sum of all forces acting upon that mass point. These forces are: • The external force 𝑓𝑖 (𝑡). • The elastic force from the neighbouring springs, negative proportional to the relative displacement. 𝑘𝑖 (𝑥𝑖−1 − 𝑥𝑖 ) + 𝑘𝑖+1 (𝑥𝑖+1 − 𝑥𝑖 ),
𝑖 = 1, . . . , 𝑛,
(1.7)
where 𝑘𝑗 is the 𝑗th spring stiffness. • The inner damping force from the neighbouring dampers, negative proportional to the relative velocity 𝑐𝑖 (𝑥˙ 𝑖−1 − 𝑥˙ 𝑖 ) + 𝑐𝑖+1 (𝑥˙ 𝑖+1 − 𝑥˙ 𝑖 ),
𝑖 = 1, . . . , 𝑛.
(1.8)
• The external damping force, negative proportional to the velocity −𝑑𝑖 𝑥˙ 𝑖 , where 𝑑𝑗 , 𝑐𝑗 are the respective inner and external damper viscosities. Here to simplify the notation we have set 𝑥0 = 𝑥𝑛+1 = 0, 𝑥˙ 0 = 𝑥˙ 𝑛+1 = 0, these are the fixed end points. Altogether we obtain (1.1) with 𝑀, 𝐶, 𝐾, from (1.2)–(1.4). All these matrices are obviously real and symmetric. By 𝑥𝑇 𝑀 𝑥 =
𝑛 ∑
𝑚𝑗 𝑥2𝑗
(1.9)
𝑗=1
and 𝑥𝑇 𝐾𝑥 = 𝑘1 𝑥21 +
𝑛 ∑
𝑘𝑗 (𝑥𝑗 − 𝑥𝑗−1 )2 + 𝑘𝑛+1 𝑥2𝑛
(1.10)
𝑗=2
both 𝑀 and 𝐾 are positive definite. By the same argument 𝐶𝑖𝑛 and 𝐶𝑜𝑢𝑡 are positive semidefinite. Exercise 1.2 How many coefficients 𝑘𝑖 in (1.3) may vanish while keeping 𝐾 positive definite? Interpret the response physically! Example 1.3 A typical external force stems from a given movement of the frame in which the vibrating structure is anchored (this is the so-called
4
1 The Model
inertial force). On the model example in Fig. 1.1 the system is anchored on two points. The force caused by the horizontal movement of these two points is taken into account by replacing the zero values of 𝑥0 , 𝑥𝑛+1 in (1.7) and (1.8) by given functions 𝑥0 (𝑡), 𝑥𝑛+1 (𝑡), respectively. This does not change the matrices 𝑀, 𝐶, 𝐾 whereas 𝑓 (𝑡) reads ⎡ ⎢ ⎢ ⎢ 𝑓 (𝑡) = ⎢ ⎢ ⎣
𝑘1 𝑥0 (𝑡) + 𝑐1 𝑥˙ 0 (𝑡) 0 .. . 0 𝑘𝑛+1 𝑥𝑛+1 (𝑡) + 𝑐𝑛+1 𝑥˙ 𝑛+1 (𝑡)
⎤ ⎥ ⎥ ⎥ ⎥. ⎥ ⎦
Exercise 1.4 Find the equilibrium configuration (i.e. the displacement 𝑥 at rest) of our model example if the external force 𝑓 (i) vanishes or (ii) is a constant vector. Hint: the matrix 𝐾 can be explicitly inverted. The model in Example 1.1 is important for other reasons too. It describes a possible discretisation of a continuous damped system (vibrating string). There the parameters 𝑐𝑖 come from the inner friction within the material whereas 𝑑𝑖 describe the external damping caused by the medium in which the system is moving (air, water) or just artificial devices (dashpots) purposely built in to calm down dangerous vibrations. In the latter case there usually will be few such dashpots resulting in the low rank matrix 𝐶𝑜𝑢𝑡 . It should be mentioned that determining the inner damping matrix for complex vibrating systems in real life may be quite a difficult task, it involves special mathematical methods as well as experimental work. From the derivations above we see that determining the equilibrium precedes any work on oscillations. The equilibrium is the first approximation to the true behaviour of the system, it is found by solving a linear or nonlinear system of equations. The next approximation, giving more detailed information are the linear oscillations around the equilibrium, that is, their movement is governed by a system of linear differential equations. Their linearity will be seen to be due to the assumption that the oscillations have small amplitudes.
1.2
Work and Energy
We will now introduce some further relevant physical notions based on Example 1.1. The work performed by any force 𝜙𝑗 (𝑡) on the 𝑗th point mass in the time interval 𝑡1 ≤ 𝑡 ≤ 𝑡2 is given by ∫
𝑡2
𝑡1
𝜙𝑗 (𝑡)𝑥˙ 𝑗 𝑑𝑡,
1.2 Work and Energy
5
this corresponds to the rule ‘work equals force times distance’. So is the work of the external force: ∫ 𝑡2
𝑡1
𝑓𝑗 (𝑡)𝑥˙ 𝑗 𝑑𝑡,
(1.11)
of the damping force: ∫
𝑡2
𝑡1
∫ [𝑐𝑗 (𝑥˙ 𝑗−1 − 𝑥˙ 𝑗 ) + 𝑐𝑗+1 (𝑥˙ 𝑗+1 − 𝑥˙ 𝑗 )]𝑥˙ 𝑗 𝑑𝑡 −
of the elastic force: ∫ 𝑡2 𝑡1
𝑡2 𝑡1
𝑑𝑗 𝑥˙ 2𝑗 𝑑𝑡,
[𝑘𝑗 (𝑥𝑗−1 − 𝑥𝑗 ) + 𝑘𝑗+1 (𝑥𝑗+1 − 𝑥𝑗 )]𝑥˙ 𝑗 𝑑𝑡.
To this we add the work of the so-called ‘inertial force’ which is given by ∫
𝑡2 𝑡1
𝑚𝑗 𝑥 ¨𝑗 𝑥˙ 𝑗 𝑑𝑡
(1.12)
The total work is the sum over all mass points, so from (1.11) and (1.12) we obtain ∫ 𝑡2 𝑥˙ 𝑇 𝑓 (𝑡)𝑑𝑡, (1.13) 𝑡1
∫
− ∫ − ∫
𝑡2
𝑡1 𝑡2 𝑡1
𝑡2
𝑡1
𝑥˙ 𝑇 𝐶 𝑥𝑑𝑡, ˙
𝑥˙ 𝑇 𝐾𝑥𝑑𝑡 = 2(𝐸𝑝 (𝑥(𝑡1 )) − 𝐸𝑝 (𝑥(𝑡2 ))),
𝑥˙ 𝑇 𝑀 𝑥 ¨𝑑𝑡 = 2(𝐸𝑘 (𝑥(𝑡2 )) − 𝐸𝑘 (𝑥(𝑡1 ))),
(1.14)
as the total work of the external forces, damping forces, elastic forces and inertial forces, respectively; here 𝐸𝑝 (𝑥) =
1 𝑇 𝑥 𝐾𝑥, 2
1 𝑇 𝑥˙ 𝑀 𝑥˙ 2 are the potential and the kinetic energy and 𝐸𝑘 (𝑥) ˙ =
𝐸(𝑥, 𝑥) ˙ = 𝐸𝑝 (𝑥) + 𝐸𝑘 (𝑥) ˙
(1.15) (1.16)
(1.17)
6
1 The Model
the total energy. Note the difference: in the first two cases the work depends on the whole motion 𝑥(𝑡), 𝑡1 ≤ 𝑡 ≤ 𝑡2 whereas in the second two cases it depends just on the values of 𝐸𝑝 , 𝐸𝑘 , respectively, taken at the end points of the motion. In the formulae (1.9), (1.10), (1.13) and (1.14) we observe a property of both potential and kinetic energy: they are ‘additive’ magnitudes, that is, the kinetic energy is a sum of the kinetic energies of single point masses whereas the potential energy is a sum of the potential energies of single springs. The relations (1.13)–(1.17) make sense and will be used for general damped systems. By premultiplying (1.1) by 𝑥˙ 𝑇 we obtain the differential energy balance 𝑑 𝐸(𝑥, 𝑥) ˙ + 𝑥˙ 𝑇 𝐶 𝑥˙ = 𝑥˙ 𝑇 𝑓 (𝑡). (1.18) 𝑑𝑡 This is obviously equivalent to the integral energy balance 𝑡2 𝐸(𝑥, 𝑥) ˙ 𝑡1 = −
∫
𝑡2
𝑡1
𝑥˙ 𝑇 𝐶 𝑥𝑑𝑡 ˙ +
∫
𝑡2
𝑡1
𝑥˙ 𝑇 𝑓 (𝑡)𝑑𝑡.
(1.19)
for any 𝑡1 ≤ 𝑡 ≤ 𝑡2 . We see that the work of the damping forces is always non-positive. This effect is called the energy dissipation. Thus, (1.19) displays the amount of energy which is transformed into thermal energy. If 𝑓 ≡ 0 then this energy loss is measured by the decrease of the total energy. If both 𝑓 ≡ 0 and 𝐶 = 0 then the total energy is preserved in time; such systems are called conservative.
1.3
The Formalism of Lagrange
The next (and last) step in our short presentation of the dynamics principles is to derive the Lagrangian formalism which is a powerful and simple tool in modelling mechanical systems. The position (also called configuration) of a mechanical system is described by the generalised coordinates 𝑞1 , . . . , 𝑞𝑠 as components of the vector 𝑞 from some region of ℝ𝑠 which is called the configuration space. The term ‘generalised’ just means that 𝑞𝑖 need not be a Euclidian coordinate as in (1.1) but possibly some other measure of position (angle and the like). To this we add one more copy of ℝ𝑠 whose elements are the generalised velocities 𝑞. ˙ The dynamical properties are described by the following four functions, defined on ℝ2𝑠 : • The kinetic energy 𝑇 = 𝑇 (𝑞, 𝑞), ˙ having a minimum equal to zero at (𝑞, 0) for any 𝑞 and with 𝑇𝑞˙′′ (𝑞, 𝑞) ˙ positive semidefinite for any 𝑞. • The potential energy 𝑉 = 𝑉 (𝑞) with 𝑉 ′ (𝑞0 ) = 0 and 𝑉 ′′ (𝑞) everywhere positive definite (that is, 𝑉 is assumed to be uniformly convex).
1.3 The Formalism of Lagrange
7
• The dissipation function 𝑄 = 𝑄(𝑞, 𝑞), ˙ having a minimum equal to zero at (𝑞0 , 0), with 𝑄′′𝑞˙ (𝑞, 𝑞) ˙ positive semidefinite for any 𝑞 as well as. • The time dependent generalised external force 𝑓 = 𝑓 (𝑡) ∈ ℝ𝑠 . Here the first three functions are assumed to be smooth enough (at least twice continuously differentiable) and the symbols ′ and ′′ denote the vector of first derivatives (gradient) and the matrix of second derivatives (Hessian), respectively. The subscript (if needed) denotes the set of variables with respect to which the derivative is taken. From these functions we construct the Lagrangian 𝐿 = 𝑇 − 𝑉 + 𝑓 𝑇 𝑞. The time development of this system is described by the Lagrange equations 𝑑 ′ 𝐿 − 𝐿′𝑞 + 𝑄′𝑞˙ = 0. 𝑑𝑡 𝑞˙
(1.20)
This is a system of differential equations of at most second order in time. Its solution, under our, rather general, conditions is not easily described. If we suppose that the solution 𝑞 = 𝑞(𝑡) does not depend on time, i.e. the system is at rest then 𝑞˙ ≡ 0 in (1.20) yields − 𝑉 ′ (𝑞) = 𝑓 with the unique solution 𝑞 = 𝑞ˆ0 (note that here 𝑓 must also be constant in 𝑡 and the uniqueness is a consequence of the uniform convexity of 𝑉 ). The point 𝑞ˆ0 is called the point of equilibrium of the system. Typical such situations are those in which 𝑓 vanishes; then 𝑞ˆ0 = 𝑞0 which is the minimum point of the potential energy. The energy balance in this general case is completely analogous to the one in (1.18) and (1.19) and its derivation is left to the reader. Remark 1.5 If 𝑓 is constant in time then we can modify the potential energy function into 𝑉ˆ (𝑞) = 𝑉 (𝑞) − 𝑞 𝑇 𝑓 thus obtaining a system with the vanishing generalised external force. If, in addition, 𝑄 = 0 then the system is conservative. This shows that there is some arbitrariness in the choice of the potential energy function. Exercise 1.6 Using the expressions (1.16), (1.15) for the kinetic and the potential energy, respectively, choose the dissipation function and the generalised external force such that, starting from (1.20) the system (1.1) is obtained. Solution. As the dissipation function take 𝑄=
1 𝑇 𝑥˙ 𝐶 𝑥; ˙ 2
8
1 The Model
then 𝐿=
1 𝑇 1 𝑥˙ 𝑀 𝑥˙ − 𝑥𝑇 𝐾𝑥 + 𝑓 𝑇 𝑥, 2 2
𝐿′𝑥˙ = 𝑀,
𝐿′𝑥 = −𝐾𝑥 + 𝑓,
𝑄′𝑥˙ = 𝐶
and (1.20) immediately yields (1.1). Now consider small oscillations around 𝑞0 , that is, the vectors 𝑞 − 𝑞0 and 𝑞˙ will be considered small for all times. This will be expected, if the distance from the equilibrium together with the velocity, taken at some initial time as well as the external force 𝑓 (𝑡) at all times are small enough. This assumed, we will approximate the functions 𝑇, 𝑉, 𝑄 by their Taylor polynomials of second order around the point (𝑞0 , 0): 1 𝑉 (𝑞) ≈ 𝑉 (𝑞0 ) + (𝑞 − 𝑞0 )𝑇 𝑉 ′ (𝑞0 ) + (𝑞 − 𝑞0 )𝑇 𝑉 ′′ (𝑞0 )(𝑞 − 𝑞0 ), 2 1 𝑇 (𝑞, 𝑞) ˙ ≈ 𝑇 (𝑞0 , 0) + 𝑞˙𝑇 𝑇𝑞˙′ (𝑞0 , 0) + 𝑞˙𝑇 𝑇𝑞′′˙ (𝑞0 , 0)𝑞˙ 2 and similarly for 𝑄. Note that, due to the conditions on 𝑇 the matrix 𝑇 ′′ (𝑞0 , 0) looks like [ ] 0 0 0 𝑇𝑞′′˙ (𝑞0 , 0) (the same for 𝑄′′ (𝑞0 , 0)). Using the properties listed above we have 𝑇 (𝑞0 , 0) = 0,
𝑇𝑞˙′ (𝑞0 , 0) = 0,
𝑄(𝑞0 , 0) = 0,
𝑄′𝑞˙ (𝑞0 , 0) = 0,
𝑉 ′ (𝑞0 ) = 0. With this approximation the Lagrange equations (1.20) yield (1.1) with 𝑥 = 𝑞 and 𝑀 = 𝑇𝑞′′˙ (𝑞0 , 0), 𝐶 = 𝑄′′𝑞˙ (𝑞0 , 0), 𝐾 = 𝑉 ′′ (𝑞0 ). (1.21) This approximation is called linearisation,1 because non-linear equations are approximated by linear ones via Taylor expansion. Under our conditions all three matrices in (1.21) are real and symmetric. In addition, 𝐾 is positive definite while the other two matrices are positive semidefinite. If we additionally assume that 𝑇 is also uniformly convex then 𝑀 will be positive definite as well. We have purposely allowed 𝑀 to be only positive semidefinite as we will analyse some cases of this kind later. Example 1.7 Consider a hydraulic model which describes the oscillation of an incompressible homogeneous heavy fluid (e.g. water) in an open vessel.
1 The
term ‘linearisation’ will be used later in a very different sense.
1.3 The Formalism of Lagrange
h1
9
s
s1
s2 h2
h D1
D2
S1
S2
Fig. 1.2 Open vessel
The vessel consists of coupled thin tubes which, for simplicity, are assumed as cylindrical (see Fig. 1.2). The parameters describing the vessel and the fluid are • ℎ1 , ℎ, ℎ2 : The fluid heights in the vertical tubes, • 𝑠1 , 𝑠, 𝑠2 : The cross sections of the vertical tubes, – 𝐷1 , 𝐷2 : The lengths, – 𝑆1 , 𝑆2 : The cross sections, – 𝑣1 , 𝑣2 : The velocities (in the sense of the arrows), in the horizontal tubes. • 𝜌: The mass density of the fluid While the values ℎ1 , ℎ, ℎ2 , 𝑣1 , 𝑣2 are functions of time 𝑡 all others are constant. Since the fluid is incompressible the volume conservation laws give 𝑠1 ℎ˙ 1 = 𝑆1 𝑣1
(1.22)
𝑆1 𝑣1 + 𝑠ℎ˙ + 𝑆2 𝑣2 = 0
(1.23)
𝑠2 ℎ˙ 2 = 𝑆2 𝑣2
(1.24)
𝑠1 ℎ˙ 1 + 𝑠ℎ˙ + 𝑠2 ℎ˙ 2 = 0
(1.25)
𝑠1 ℎ1 + 𝑠ℎ + 𝑠2 ℎ2 = 𝛤,
(1.26)
These equalities imply
which, integrated, gives
where 𝛤 is a constant (in fact, 𝛤 + 𝐷1 𝑆1 + 𝐷2 𝑆2 is the fixed total volume of the fluid). Thus, the movement of the system is described by ℎ1 = ℎ1 (𝑡), ℎ2 = ℎ2 (𝑡).
10
1 The Model
To obtain the Lagrange function we must find the expressions for the kinetic and the potential energy as well as for a dissipation function – as functions of ℎ1 , ℎ2 , ℎ˙ 1 , ℎ˙ 2 . In each vertical tube the potential energy equals the mass times half the height (this is the center of gravity) times 𝑔, the gravity acceleration. The total potential energy 𝑉 is given by 𝑉 𝑉 (ℎ1 , ℎ2 ) ℎ2 ℎ2 ℎ2 = = 𝑠1 1 + 𝑠 + 𝑠2 2 𝜌𝑔 𝜌𝑔 2 2 2
(1.27)
ℎ21 (𝛤 − 𝑠1 ℎ1 − 𝑠2 ℎ2 )2 ℎ2 + + 𝑠2 2 2 2𝑠 2
(1.28)
= 𝑠1
(the contributions from the horizontal tubes are ignored since they do not depend on ℎ1 , ℎ2 , ℎ˙ 1 , ℎ˙ 2 and hence do not enter the Lagrange equations). It is readily seen that 𝑉 takes its minimum at ˆ := ℎ1 = ℎ = ℎ2 = ℎ
𝛤 . 𝑠1 + 𝑠 + 𝑠2
(1.29)
The kinetic energy in each tube equals half the mass times the square of the velocity. The total kinetic energy 𝑇 is given by 𝑇 /𝜌 = 𝑇 (ℎ1 , ℎ2 , ℎ˙ 1 , ℎ˙ 2 )/𝜌
= ℎ 1 𝑠1
=
ℎ˙ 21 𝑣2 ℎ˙ 2 𝑣2 ℎ˙ 2 + 𝐷1 𝑆1 1 + ℎ𝑠 + 𝐷2 𝑆2 2 + ℎ2 𝑠2 2 2 2 2 2 2
(1.30)
( ) ( ) 𝐷1 𝑠21 ℎ˙ 21 (𝑠1 ℎ˙ 1 + 𝑠2 ℎ˙ 2 )2 𝐷2 𝑠22 ℎ˙ 22 ℎ1 𝑠1 + + (𝛤 − 𝑠1 ℎ1 − 𝑠2 ℎ2 ) + ℎ 𝑠 + , 2 2 𝑆1 2 2𝑠2 𝑆2 2
(1.31)
where ℎ˙ was eliminated by means of (1.25). While 𝑇 and 𝑉 result almost canonically from first principles and the geometry of the system the dissipation function allows more freedom. We will assume that the frictional force in each tube is proportional to the velocity of the fluid and to the fluid-filled length of the tube. This leads to the dissipation function 𝑄 = 𝑄(ℎ1 , ℎ2 , ℎ˙ 1 , ℎ˙ 2 )
= ℎ 1 𝜃1
ℎ˙ 21 𝑣2 ℎ˙ 2 𝑣2 ℎ˙ 2 + 𝐷1 𝛩1 1 + ℎ𝜃 + 𝐷2 𝛩2 2 + ℎ2 𝜃2 2 2 2 2 2 2
(1.32)
1.3 The Formalism of Lagrange
=
11
( ) 𝐷1 𝛩1 𝑠21 ℎ˙ 21 (𝑠1 ℎ˙ 1 + 𝑠2 ℎ˙ 2 )2 ℎ1 𝜃 1 + + (𝛤 − 𝑠 ℎ − 𝑠 ℎ ) 𝛩 1 1 2 2 𝑆12 2 2𝑠3 ( ) 𝐷2 𝛩2 𝑠22 ℎ˙ 22 + ℎ 2 𝜃2 + , 𝑆22 2
(1.33) (1.34)
where 𝜃1 , 𝛩1 , 𝜃, 𝛩2 , 𝜃2 are positive constants characterising the friction per unit length in the respective tubes. By inserting the obtained 𝑇, 𝑉, 𝑄 into [ ]𝑇 (1.20) and by taking 𝑞 as ℎ1 ℎ2 we obtain two non-linear differential equations of second order for the unknown functions ℎ1 , ℎ2 . Now comes the linearisation at the equilibrium position which is given by (1.29). The Taylor expansion of second order for 𝑇, 𝑉, 𝑄 around the point ˆ ℎ˙ 1 = ℎ˙ 2 = 0 takes into account that all three gradients at that ℎ1 = ℎ2 = ℎ, point vanish thus giving ( ) 𝜒21 (𝑠1 𝜒1 + 𝑠2 𝜒2 )2 𝜒22 ˆ ˆ 𝑉 ≈ 𝑉 (ℎ, ℎ) + 𝜌𝑔 𝑠1 + + 𝑠2 2 2𝑠 2 1 ˆ ˆ = 𝑉 (ℎ, ℎ) + 𝜒𝑇 𝐾𝜒 2
(1.35)
[ ] ˆ ℎ2 − ℎ ˆ 𝑇 and with 𝜒 = ℎ1 − ℎ (( 𝑇 ≈𝜌
2 ˆℎ𝑠1 + 𝐷1 𝑠1 𝑆1
)
( ) 2) 2 𝜒˙ 21 ˆ (𝑠1 𝜒˙ 1 + 𝑠2 𝜒˙ 2 )2 ˆ 2 + 𝐷2 𝑠2 𝜒˙ 2 +ℎ + ℎ𝑠 2 2𝑠 𝑆2 2 (1.36) =
( 𝑄≈
2 ˆℎ𝜃1 + 𝐷1 𝛩1 𝑠1 𝑆12
)
1 𝑇 𝜒˙ 𝑀 𝜒˙ 2 (1.37)
( ) 𝜒˙ 21 ˆ (𝑠1 𝜒˙ 1 + 𝑠𝑎2 𝜒˙ 2 )2 𝐷2 𝛩2 𝑠22 𝜒˙ 22 ˆ +ℎ 𝛩 + ℎ𝜃2 + 2 2𝑠2 𝑆22 2 (1.38) =
1 𝑇 𝜒˙ 𝐶 𝜒. ˙ 2 (1.39)
The symmetric matrices 𝑀, 𝐶, 𝐾 are immediately recoverable from (1.35)– (1.38) and are positive definite by their very definition. By inserting these 𝑇, 𝑉, 𝑄 into (1.20) the equations (1.1) result with 𝑓 = 0.
12
1 The Model
Any linearised model can be assessed in two ways: • By estimating the error between the true model and the linearised one and • By investigating the well-posedness of the linear model itself: do small forces 𝑓 (𝑡) and small initial data produce small solutions? While the first task is completely out of the scope of the present text, the second one will be given our attention in the course of these Notes.
1.4
Oscillating Electrical Circuits
Equation (1.1) governs some other physical systems, most relevant among them is the electrical circuit system which we will present in the following example. Example 1.8 𝑛 simple wire loops (circuits) are placed in a chain. The 𝑗th circuit is characterised by the current 𝐼𝑗 , the impressed electromagnetic force 𝐸𝑗 (𝑡), the capacity 𝐶𝑗 > 0, the resistance 𝑅𝑗 ≥ 0, and the left and the right inductance 𝐿𝑗 > 0, 𝑀𝑗 > 0, respectively. In addition, the neighbouring 𝑗th and 𝑗 + 1th inductances are accompanied by the mutual inductance 𝑁𝑗 ≥ 0. The circuits are shown in Fig. 1.3 with 𝑛 = 3. The inductances satisfy the inequality √ 𝑁𝑗 < 𝑀𝑗 𝐿𝑗+1 . The equation governing the time behaviour of this system is derived from the fundamental electromagnetic laws and it reads in vector form ˙ ℒ𝐼¨ + ℛ𝐼˙ + 𝒦𝐼 = 𝐸(𝑡). Here
⎡ ⎢ ⎢ ⎢ ⎢ ℒ=⎢ ⎢ ⎢ ⎣
𝐿1 + 𝑀1 𝑁1
𝑁1
⎤
⎥ . ⎥ 𝐿2 + 𝑀2 𝑁2 . . ⎥ ⎥ .. .. .. ⎥ . . . ⎥ ⎥ .. .. . . 𝑁𝑛−1 ⎦ 𝑁𝑛−1 𝐿𝑛 + 𝑀𝑛
is the inductance matrix, ℛ = diag(𝑅1 , . . . , 𝑅𝑛 ) is the resistance matrix and 𝒦 = diag(1/𝐶1 , . . . , 1/𝐶𝑛 )
1.4 Oscillating Electrical Circuits
13
E1(t)
L1
E2(t)
M1
E3(t)
L2
M2
M3
N2
N1 C1
L3
R1
C2
R2
C3
R3
Fig. 1.3 Circuits
is the inverse capacitance matrix. The latter two matrices are obviously positive definite whereas the positive definiteness of the first one is readily deduced from its structure and the inequality (1.40). Exercise 1.9 Show the positive definiteness of the inductance matrix ℒ. Hint: use the fact that each of the matrices [
is positive definite.
𝑀𝑗 𝑁𝑗 𝑁𝑗 𝐿𝑗+1
]
Chapter 2
Simultaneous Diagonalisation (Modal Damping)
In this chapter we describe undamped and modally damped systems. They are wholly explained by the knowledge of the mass and the stiffness matrix. This is the broadly known case and we shall outline it here, not only because it is an important special case, but because it is often used as a starting position in the analysis of general damped systems.
2.1
Undamped Systems
The system (1.1) is called undamped,, if the damping vanishes: 𝐶 = 0. The solution of an undamped system is best described by the generalised eigenvalue decomposition of the matrix pair 𝐾, 𝑀 : 𝛷𝑇 𝐾𝛷 = diag(𝜇1 , . . . , 𝜇𝑛 ),
𝛷𝑇 𝑀 𝛷 = 𝐼.
(2.1)
We say that the matrix 𝛷 reduces the pair 𝐾, 𝑀 of symmetric matrices to diagonal form by congruence. This reduction is always possible, if the matrix 𝑀 is positive definite. Instead of speaking of the matrix pair one often speaks of the matrix pencil, (that is, matrix function) 𝐾 − 𝜆𝑀 . If 𝑀 = 𝐼 then (2.1) reduces to the (standard) eigenvalue decomposition valid for any symmetric matrix 𝐾, in this case the matrix 𝛷 is orthogonal. An equivalent way of writing (2.1) is 𝐾𝛷 = 𝑀 𝛷 diag(𝜇1 , . . . , 𝜇𝑛 ), or also
𝐾𝜙𝑗 = 𝜇𝑗 𝑀 𝜙𝑗 ,
𝛷𝑇 𝑀 𝛷 = 𝐼.
(2.2)
𝜙𝑇𝑗 𝑀 𝜙𝑘 = 𝛿𝑘𝑗 .
Thus, the columns 𝜙𝑗 of 𝛷 form an 𝑀 -orthonormal basis of eigenvectors of the generalised eigenvalue problem 𝐾𝜙 = 𝜇𝑀 𝜙, K. Veseli´ c, Damped Oscillations of Linear Systems, Lecture Notes in Mathematics 2023, DOI 10.1007/978-3-642-21335-9 2, © Springer-Verlag Berlin Heidelberg 2011
(2.3) 15
16
2 Simultaneous Diagonalisation (Modal Damping)
whereas 𝜇𝑘 are the zeros of the characteristic polynomial det(𝐾 − 𝜇𝑀 ) of the pair 𝐾, 𝑀 . Hence 𝜇=
𝜙𝑇 𝐾𝜙 𝜙𝑇 𝐾𝜙𝑘 , in particular, 𝜇𝑘 = 𝑇𝑘 𝑇 𝜙 𝑀𝜙 𝜙𝑘 𝑀 𝜙𝑘
shows that all 𝜇𝑘 are positive, if both 𝐾 and 𝑀 are positive definite as in our case. So we may rewrite (2.1) as 𝛷𝑇 𝐾𝛷 = 𝛺 2 , with
𝛷𝑇 𝑀 𝛷 = 𝐼
𝛺 = diag(𝜔1 , . . . , 𝜔𝑛 ),
𝜔𝑘 =
√
(2.4) 𝜇𝑘
(2.5)
The quantities 𝜔𝑘 will be called the eigenfrequencies of the system (1.1) with 𝐶 = 0. The generalised eigenvalue decomposition can be obtained by any common matrix computation package (e.g. by calling eig(K,M) in MATLAB). The solution of the homogeneous equation 𝑀𝑥 ¨ + 𝐾𝑥 = 0
(2.6)
is given by the formula ⎡
⎤ 𝑎1 cos 𝜔1 𝑡 + 𝑏1 sin 𝜔1 𝑡 ⎢ ⎥ .. 𝑥(𝑡) = 𝛷 ⎣ ⎦, .
𝑎 = 𝛷−1 𝑥0 ,
𝜔𝑘 𝑏𝑘 = (𝛷−1 𝑥˙ 0 )𝑘 ,
𝑎𝑛 cos 𝜔𝑛 𝑡 + 𝑏𝑛 sin 𝜔𝑛 𝑡
(2.7) which is readily verified. The values 𝜔𝑘 are of interest even if the damping 𝐶 does not vanish and in this context they are called the undamped frequencies of the system (1.1). In physical language the formula (2.7) is oft described by the phrase ‘any oscillation is a superposition of harmonic oscillations or eigenmodes’ which are 𝜙𝑘 (𝑎𝑘 cos 𝜔𝑘 𝑡 + 𝑏𝑘 sin 𝜔𝑘 𝑡), 𝑘 = 1, . . . , 𝑛. Exercise 2.1 Show that the eigenmodes are those solutions 𝑥(𝑡) of the equation (2.6) in which ‘all particles oscillate in the same phase’ that is, 𝑥(𝑡) = 𝑥0 𝑇 (𝑡), where 𝑥0 is a fixed non-zero vector and 𝑇 (𝑡) is a scalar-valued function of 𝑡 (the above formula is also well known under the name ‘Fourier ansatz’).
2.1 Undamped Systems
17
The eigenvalues 𝜇𝑘 , taken in the non-decreasing ordering, are given by the known minimax formula 𝜇𝑘 = max
𝑆𝑛−𝑘+1
min
𝑥∈𝑆𝑛−𝑘+1 𝑥∕=0
𝑥𝑇 𝐾𝑥 𝑥𝑇 𝐾𝑥 = min max 𝑇 , 𝑇 𝑆𝑘 𝑥∈𝑆𝑘 𝑥 𝑀 𝑥 𝑥 𝑀𝑥 𝑥∕=0
(2.8)
where 𝑆𝑗 denotes any subspace of dimension 𝑗. We will here skip proving these – fairly known – formulae, valid for any pair 𝐾, 𝑀 of symmetric matrices with 𝑀 positive definite. We will, however, provide a proof later within a more general situation (see Chap. 10 below). The eigenfrequencies have an important monotonicity property. We introduce the relation called relative stiffness in the set of all pairs of positive ˆ 𝑀 ˆ is definite symmetric matrices 𝐾, 𝑀 as follows. We say that the pair 𝐾, ˆ ˆ relatively stiffer than 𝐾, 𝑀 , if the matrices 𝐾 − 𝐾 and 𝑀 − 𝑀 are positive semidefinite (that is, if stiffness is growing and the mass is falling). Theorem 2.2 Increasing relative stiffness increases the eigenfrequencies. ˆ − 𝐾 and 𝑀 − 𝑀 ˆ are positive semidefinite then the More precisely, if 𝐾 corresponding non-decreasingly ordered eigenfrequencies satisfy 𝜔𝑘 ≤ 𝜔 ˆ𝑘. Proof. Just note that
ˆ 𝑥𝑇 𝐾𝑥 𝑥𝑇 𝐾𝑥 ≤ ˆ𝑥 𝑥𝑇 𝑀 𝑥 𝑥𝑇 𝑀 for all non-vanishing 𝑥. Then take first minimum and then maximum and the statement follows from (2.8). Q.E.D. ˆ is generated by the spring stiffnesses 𝑘ˆ𝑗 If in Example 1.1 the matrix 𝐾 ˆ − 𝐾 we have then by (1.10) for 𝛿𝐾 = 𝐾 𝑥𝑇 𝛿𝐾𝑥 = 𝛿𝑘1 𝑥21 +
𝑛 ∑
𝛿𝑘𝑗 (𝑥𝑗 − 𝑥𝑗−1 )2 + 𝛿𝑘𝑛+1 𝑥2𝑛 ,
(2.9)
𝑗=2
where
𝛿𝑘𝑗 = 𝑘ˆ𝑗 − 𝑘𝑗 .
So, 𝑘ˆ𝑗 ≥ 𝑘𝑗 implies the positive semidefiniteness of 𝛿𝐾, that is the relative ˆ − 𝑀, stiffness is growing. The same happens with the masses: take 𝛿𝑀 = 𝑀 then 𝑛 ∑ 𝑥𝑇 𝛿𝑀 𝑥 = 𝛿𝑚𝑗 𝑥2𝑗 , 𝛿𝑚𝑗 = 𝑚 ˆ 𝑗 − 𝑚𝑗 𝑗=1
18
2 Simultaneous Diagonalisation (Modal Damping)
and 𝑚 ˆ 𝑗 ≤ 𝑚𝑗 implies the negative semidefiniteness of 𝛿𝑀 – the relative stiffness is again growing. Thus, our definition of the relative stiffness has deep physical roots. The next question is: how do small changes in the system parameters 𝑘𝑗 , 𝑚𝑗 affect the eigenvalues? We make the term ‘small changes’ precise as follows ∣𝛿𝑘𝑗 ∣ ≤ 𝜖𝑘𝑗 , ∣𝛿𝑚𝑗 ∣ ≤ 𝜂𝑚𝑗 (2.10) with 0 ≤ 𝜖, 𝜂 < 1. This kind of relative error is typical both in physical measurements and in numerical computations, in fact, in floating point arithmetic 𝜖, 𝜂 ≈ 10−𝑑 where 𝑑 is the number of significant digits in a decimal number. The corresponding errors in the eigenvalues will be an immediate consequence of (2.10) and Theorem 2.2. Indeed, from (2.9) and (2.10) it follows ∣𝑥𝑇 𝛿𝐾𝑥∣ ≤ 𝜖𝑥𝑇 𝐾𝑥, Then and
∣𝑥𝑇 𝛿𝑀 𝑥∣ ≤ 𝜂𝑥𝑇 𝑀 𝑥.
(2.11)
ˆ ≤ (1 + 𝜖)𝑥𝑇 𝐾𝑥 (1 − 𝜖)𝑥𝑇 𝐾𝑥 ≤ 𝑥𝑇 𝐾𝑥 ˆ 𝑥 ≤ (1 + 𝜂)𝑥𝑇 𝑀 𝑥 (1 − 𝜂)𝑥𝑇 𝑀 𝑥 ≤ 𝑥𝑇 𝑀
such that the pairs (1 − 𝜖)𝐾, (1 + 𝜂)𝑀 ;
ˆ 𝑀 ˆ; 𝐾,
(1 + 𝜖)𝐾, (1 − 𝜂)𝑀
are ordered in growing relative stiffness. Therefore by Theorem 2.2 the corresponding eigenvalues 1−𝜖 𝜇𝑘 , 1+𝜂 satisfy
𝜇 ˆ𝑘 ,
1+𝜖 𝜇𝑘 1−𝜂
1−𝜖 1+𝜖 𝜇𝑘 ≤ 𝜇 ˆ𝑘 ≤ 𝜇𝑘 1+𝜂 1−𝜂
(2.12)
(and similarly for the respective 𝜔𝑘 , 𝜔 ˆ 𝑘 ). In particular, for 𝛿𝜇𝑘 = 𝜇 ˆ𝑘 − 𝜇𝑘 the relative error estimates 𝜖+𝜂 ∣𝛿𝜇𝑘 ∣ ≤ 𝜇𝑘 (2.13) 1−𝜂 are valid. Note that both (2.12) and (2.13) are quite general. They depend only on the bounds (2.11), the only requirement is that both matrices 𝐾, 𝑀 be symmetric and positive definite.
2.2 Frequencies as Singular Values
19
ˆ = 𝑀 = 𝐼 the more commonly known error estimate holds In the case 𝑀 ˆ − 𝐾) ≤ 𝜇 ˆ − 𝐾) 𝜇𝑘 + min 𝜎(𝐾 ˆ𝑘 ≤ 𝜇𝑘 + max 𝜎(𝐾 and in particular
ˆ − 𝐾∥. ∣𝛿𝜇𝑘 ∣ ≤ ∥𝐾
(2.14) (2.15)
The proof again goes by immediate application of Theorem 2.2 and is left to the reader.
2.2
Frequencies as Singular Values
There is another way to compute the eigenfrequencies 𝜔𝑗 . We first make the decomposition 𝐾 = 𝐿1 𝐿𝑇1 , 𝑀 = 𝐿2 𝐿𝑇2 , (2.16) 𝑦1 = 𝐿𝑇1 𝑥,
𝑦2 = 𝐿𝑇2 𝑥, ˙
(here 𝐿1 , 𝐿2 may, but need not be Cholesky factors). Then we make the singular value decomposition 𝑇 𝐿−1 2 𝐿1 = 𝑈 𝜮𝑉
(2.17)
where 𝑈, 𝑉 are real orthogonal matrices and 𝜮 is diagonal with positive diagonal elements. Hence 𝑇 −𝑇 𝐿−1 = 𝑈 𝜮 2𝑈 𝑇 2 𝐿 1 𝐿 1 𝐿2
or
𝐾𝛷 = 𝑀 𝛷𝜮 2 ,
𝛷 = 𝐿−𝑇 2 𝑈
Now we can identify this 𝛷 with the one from (2.4) and 𝜮 with 𝛺. Thus the eigenfrequencies of the undamped system are the singular values of the matrix 1 𝐿−1 2 𝐿1 . The computation of 𝛺 by (2.17) may have advantages over the one by (2.2), in particular, if 𝜔𝑗 greatly differ from each other. Indeed, by setting in Example 1.1 𝑛 = 3, 𝑘4 = 0, 𝑚𝑖 = 1 the matrix 𝐿1 is directly obtained as ⎡
⎤ 𝜅1 −𝜅2 0 𝐿1 = ⎣ 0 𝜅2 −𝜅3 ⎦ , 0 0 𝜅3
𝜅𝑖 =
√ 𝑘𝑖 .
(2.18)
If we take 𝑘1 = 𝑘2 = 1, 𝑘3 ≫ 1 (that is, the third spring is almost rigid) then the way through (2.2) may spoil the lower frequency. For instance, with 1 Equivalently
we may speak of 𝜔𝑗 as the generalised singular values of the pair 𝐿1 , 𝐿2 .
20
2 Simultaneous Diagonalisation (Modal Damping)
the value 𝑘3 = 9.999999 ⋅ 1015 the double-precision computation with Matlab gives the frequencies sqrt(eig(K,M))
svd(L_2\L_1)
7.962252170181258e-01 1.538189001320851e+00 1.414213491662415e+08
4.682131924621356e-01 1.510223959022110e+00 1.414213491662415e+08
The singular value decomposition gives largely correct low eigenfrequencies. This phenomenon is independent of the eigenvalue or singular value algorithm used and it has to do with the fact that standard eigensolution algorithms compute the lowest eigenvalue of (2.2) with the relative error ≈ 𝜖𝜅(𝐾𝑀 −1 ), that is, the machine precision 𝜖 is amplified by the con−1 16 dition number 𝜅(𝐾𝑀 √ ) ≈ 10 whereas the same error with (2.17) is −1 −1 ≈ 𝜖𝜅(𝐿2 𝐿1 ) = 𝜖 𝜅(𝐾𝑀 ) (cf. e.g. [19]). In the second case the amplification is the square root of the first one!
2.3
Modally Damped Systems
Here we study those damped systems which can be completely explained by their undamped part. In order to do this it is convenient to make a coordinate transformation; we set 𝑥 = 𝛷𝑥′ , (2.19) where 𝛷 is any real non-singular matrix. Thus (1.1) goes over into 𝑀 ′𝑥 ¨′ + 𝐶 ′ 𝑥˙ ′ + 𝐾 ′ 𝑥′ = 𝑔(𝑡),
(2.20)
with 𝑀 ′ = 𝛷𝑇 𝑀 𝛷,
𝐶 ′ = 𝛷𝑇 𝐶𝛷,
𝐾 ′ = 𝛷𝑇 𝐾𝛷,
𝑔 = 𝛷𝑇 𝑓.
(2.21)
Choose now the matrix 𝛷 as in the previous section, that is, 𝛷𝑇 𝑀 𝛷 = 𝐼,
𝛷𝑇 𝐾𝛷 = 𝛺 = diag(𝜔12 , . . . , 𝜔𝑛2 ).
(The right hand side 𝑓 (𝑡) in (2.20) can always be taken into account by the Duhamel’s term as in (3.1) so we will mostly restrict ourselves to consider 𝑓 = 0 which corresponds to a ‘freely oscillating’ system.) Now, if 𝐷 = (𝑑𝑗𝑘 ) = 𝛷𝑇 𝐶𝛷 (2.22)
2.3 Modally Damped Systems
21
is diagonal as well then (1.1) is equivalent to 𝜉¨𝑘 + 𝑑𝑘𝑘 𝜉˙𝑘 + 𝜔𝑘2 𝜉𝑘 = 0,
𝑥 = 𝛷𝜉
with the known solution 𝜉𝑘 = 𝑎𝑘 𝑢+ (𝑡, 𝜔𝑘 , 𝑑𝑘𝑘 ) + 𝑏𝑘 𝑢− (𝑡, 𝜔𝑘 , 𝑑𝑘𝑘 ),
(2.23)
+
𝑢+ (𝑡, 𝜔, 𝑑) = 𝑒𝜆 (𝜔,𝑑) 𝑡 , { − 𝑒𝜆 (𝜔,𝑑) 𝑡 , 𝛿(𝜔, 𝑑) ∕= 0 − 𝑢 (𝑡, 𝜔, 𝑑) = + 𝑡 𝑒𝜆 (𝜔,𝑑) 𝑡 , 𝛿(𝜔, 𝑑) = 0 where 𝛿(𝜔, 𝑑) = 𝑑2 − 4𝜔 2 , √ −𝑑 ± 𝛿(𝜔, 𝑑) ± 𝜆 (𝜔, 𝑑) = . 2 The constants 𝑎𝑘 , 𝑏𝑘 in (2.23) are obtained from the initial data 𝑥0 , 𝑥˙ 0 similarly as in (2.7). However, the simultaneous diagonalisability of the three matrices 𝑀, 𝐶, 𝐾 is rather an exception, as shown by the following theorem. Theorem 2.3 Let 𝑀, 𝐶, 𝐾 be as in (1.1). A non-singular 𝛷 such that the matrices 𝛷𝑇 𝑀 𝛷, 𝛷𝑇 𝐶𝛷, 𝛷𝑇 𝐾𝛷 are diagonal exists, if and only if 𝐶𝐾 −1 𝑀 = 𝑀 𝐾 −1 𝐶.
(2.24)
Proof. The ‘only if part’ is trivial. Conversely, using (2.4) the identity (2.24) yields 𝐶𝛷𝛺 −2 𝛷−1 = 𝛷−𝑇 𝛺 −2 𝛷𝑇 𝐶, hence and then
𝛷𝑇 𝐶𝛷𝛺 −2 = 𝛺 −2 𝛷𝑇 𝐶𝛷 𝛺 2 𝛷𝑇 𝐶𝛷 = 𝛷𝑇 𝐶𝛷𝛺 2
i.e. the two real symmetric matrices 𝛺 2 and 𝐷 = 𝛷𝑇 𝐶𝛷 commute and 𝛺 2 is diagonal, so there exists a real orthogonal matrix 𝑈 such that 𝑈 𝑇 𝛺 2 𝑈 = 𝛺 2 , and 𝑈 𝑇 𝐷𝑈 = diag(𝑑11 , . . . , 𝑑𝑛𝑛 ). Indeed, since the diagonal elements of 𝛺 are non-decreasingly ordered we may write 𝛺 = diag(𝛺1 , . . . , 𝛺𝑝 ),
22
2 Simultaneous Diagonalisation (Modal Damping)
where 𝛺1 , . . . , 𝛺𝑝 are scalar matrices corresponding to distinct spectral points of 𝛺. Now, 𝐷𝛺 2 = 𝛺 2 𝐷 implies 𝐷𝛺 = 𝛺𝐷 and therefore 𝐷 = diag(𝐷1 , . . . , 𝐷𝑝 ), with the same block partition. The matrices 𝐷1 , . . . , 𝐷𝑝 are real symmetric, so there are orthogonal matrices 𝑈1 , . . . , 𝑈𝑝 such that all 𝑈𝑗𝑇 𝐷𝑗 𝑈𝑗 are diagonal. By setting 𝛷1 = 𝛷 diag(𝐷1 , . . . , 𝐷𝑝 ) all three matrices 𝛷𝑇1 𝑀 𝛷1 , 𝛷𝑇1 𝐶𝛷1 , 𝛷𝑇1 𝐾𝛷1 are diagonal. Q.E.D. Exercise 2.4 Show that Theorem 2.3 remains valid, if 𝑀 is allowed to be only positive semidefinite. Exercise 2.5 Prove that (2.24) holds, if 𝛼𝑀 + 𝛽𝐶 + 𝛾𝐾 = 0, where not all of 𝛼, 𝛽, 𝛾 vanish (proportional damping). When is this the case with 𝐶 from (1.4)–(1.6)? Exercise 2.6 Prove that (2.24) is equivalent to 𝐶𝑀 −1 𝐾 = 𝐾𝑀 −1 𝐶 and also to 𝐾𝐶 −1 𝑀 = 𝑀 𝐶 −1 𝐾, provided that these inverses exist. Exercise 2.7 Try to find a necessary and sufficient condition that 𝐶 from (1.4)–(1.6) satisfies (2.24). If 𝐶 satisfies (2.24) then we say that the system (1.1) is modally damped.
Chapter 3
Phase Space
In general, simultaneous diagonalisation of all three matrices M, C, K is not possible and the usual transformation to a system of first order is performed. This transformation opens the possibility of using powerful means of the spectral theory in order to understand the time behaviour of damped systems.
3.1
General Solution
If the matrix M is non-singular the existence and uniqueness of the solution of the initial value problem M x ¨ + C x˙ + Kx = f (t) with the standard initial conditions x(0) = x0 , x(0) ˙ = x˙ 0 is insured by the standard theory for linear systems of differential equations which we now review. We rewrite (1.1) as x ¨ = −M −1Kx − M −1 C x˙ + M −1 f (t) and by setting x1 = x, x2 = x˙ as d x1 x = B 1 + g(t), dt x2 x2 0 1 , B= −M −1 K −M −1 C
0 g(t) = . M −1 f (t)
K. Veseli´ c, Damped Oscillations of Linear Systems, Lecture Notes in Mathematics 2023, DOI 10.1007/978-3-642-21335-9 3, © Springer-Verlag Berlin Heidelberg 2011
23
24
3 Phase Space
This equation is solved by the so-called Duhamel formula
with eBt =
x1 x2
= eBt
∞ B j tj
j!
j=0
t x0 + eB(t−τ ) g(τ ) dτ x˙ 0 0
(3.1)
∞
d Bt B j tj−1 e = = BeBt . dt (j − 1)! j=1
,
For the uniqueness: if x, y solve (1.1) then u = y − x, u1 = u, u2 = u˙ satisfy d dt
u1 u2
=B
By setting
we obtain
d dt
v1 v2
u1 , u2
u1 (0) = u2 (0) = 0.
v1 u = e−Bt 1 v2 u2
= 0,
v1 (0) = v2 (0) = 0
−1 = eBt also u1 (t) ≡ u2 (t) ≡ 0. so, v1 (t) ≡ v2 (t) ≡ 0 and by e−Bt
3.2
Energy Phase Space
The shortcoming of the transformation in Chap. 3.1 is that in the system matrix the symmetry properties of the matrices M, C, K are lost. A more careful approach will try to keep as much structure as possible and this is what we will do in this chapter. An important feature of this approach is that the Euclidian norm on the underlying space (called phase space) is closely related to the total energy of the system. We start with any decomposition (2.16). By substituting y1 = LT1 y, y2 = T L2 y˙ (1.1) becomes y˙ = Ay + g(t), A= which is solved by
g(t) =
0 , L−1 2 f (t)
LT1 L−T 2 −1 −T −L−1 2 L1 −L2 CL2 0
(3.2)
(3.3)
3.2 Energy Phase Space
25
y=e
At
t y10 eA(t−τ ) g(τ ) dτ, + y20 0
y10 = LT1 x0 ,
(3.4)
y20 = LT2 x˙ 0 .
The differential equation (3.2) will be referred to as the evolution equation. The formula (3.4) is, of course, equivalent to (3.1). The advantage of (3.4) is due to the richer structure of the phase-space matrix A in (3.3). It has two important properties: −1 T yT Ay = −(L−1 2 y2 ) CL2 y2 ≤ 0 for all y,
(3.5)
AT = JAJ,
(3.6)
where J=
I 0 . 0 −I
(3.7)
The property (3.6) is called J-symmetry, a general real J-symmetric matrix A has the form A11 A12 , A= −AT12 A22
AT11 = A11 ,
AT22 = A22 ,
so, A is ‘symmetric up to signs’. Moreover, for y = y(t) satisfying (1.1) y2 = y1 2 + y2 2 = xT Kx + x˙ T M x. ˙
(3.8)
˙ its kinetic Since xT Kx/2 is the potential energy of the system and x˙ T M x/2 energy we see that y2 is twice the total energy of the system (1.1). That is why we call the representation (3.2), (3.3) of the system (1.1) the energy phase space representation. For g(t) ≡ 0 (freely oscillating damped system) we have by (3.5) d d y2 = y T y = y˙ T y + y T y˙ = y T (A + AT )y = 2y T Ay ≤ 0 dt dt
(3.9)
that is, the energy of such a system is a non-increasing function of time. If C = 0 then the energy is constant in time. The property (3.5) of the matrix A is called dissipativity. Exercise 3.1 Show that the dissipativity of A is equivalent to the contractivity of the exponential: eAt ≤ 1,
t ≥ 0.
26
3 Phase Space
Remark 3.2 There are infinitely many decompositions (2.16). Most common are • The Cholesky decomposition. choice. • L1 = Φ−T Ω, L2 = Φ−T , where and 0 A= −Ω
This is numerically the least expensive Φ, Ω are given by (2.4). Then L−1 2 L1 = Ω Ω (3.10) , D = ΦT CΦ. −D
This is the modal representation of the phase-space matrix. √ √ • The positive definite square roots: L1 = K, L2 = M . but some others will also prove useful. Whenever no confusion is expected the matrix D in (3.10) will also be called the damping matrix (in modal coordinates). In practice one is often given a damped system just in terms of the matrices D and Ω in the modal representation. A common damping matrix is of the form D = ρΩ + W W T , where ρ is a small parameter (a few percent) describing the inner damping of the material whereas the matrix W is injective of order n × m with m n. The term W W T describes a few damping devices (dashpots) built in order to prevent large displacements (see [17]). Any transition from (1.1) to a first order system with the substitution y1 = W1 x, y2 = W2 x˙ where W1 , W2 are non-singular matrices is called a linearisation of the system (1.1). As we have already said we use the term ‘linearisation’ in two quite different senses, but since both of them are already traditional and very different in their meaning we hope not to cause confusion by using them thus. Exercise 3.3 Prove that any two linearisations lead to phase-space matrices which are similar to each other. Investigate under what linearisations the corresponding phase-space matrix will be (i) dissipative, (ii) J-symmetric. The transformation of coordinates (2.19)–(2.21) can be made on any damped system (1.1) and with any non-singular transformation matrix Φ thus leading to (2.20). The energy expressions stay invariant: x˙ T M x˙ = x˙ T M x˙ ,
xT Kx = xT K x .
We now derive the relation between the two phase-space matrices A and A , built by (3.3) from M, C, K and M , C , K , respectively. So we take L1 , L2 from (2.16) and L1 , L2 from K = ΦT L1 LT1 Φ = L1 LT 1 ,
M = ΦT L2 LT2 Φ = L2 LT 2 .
3.2 Energy Phase Space
27
Hence the matrices U1 = LT1 ΦL−T and U2 = LT2 ΦL−T are unitary. Thus, 1 2 −1 −1 L−1 2 L1 = U2 L2 L1 U1
and
−T −T L−1 = U2−1 L−1 2 C L2 2 CL2 U2 .
Altogether
0 U1−1 LT1 L−T 2 U2 −1 −1 −T −U2−1L−1 2 L1 U1 −U2 L2 CL2 U2 −1 0 LT1 L−T U1 0 0 U1 2 = −1 −T −L−1 0 U2−1 0 U2 2 L1 −L2 CL2 A =
= U −1 AU
with
U1 0 U= 0 U2
(3.11)
which is unitary. Although no truly damped system has a normal phase-space matrix something else is true: among a large variety of possible linearisation matrices the one in (3.3) is closest to normality. We introduce the departure from normality of a general matrix A of order n as η(A) = A2E −
n
|λi |2 ,
i=1
where AE =
|aij |2
ij
is the Euclidean norm of A and λi its eigenvalues. By a unitary transformation of A to triangular form it is seen that η(A) is always non-negative. Moreover, η(A) vanishes if and only if A is normal. The departure from normality is a measure of sensitivity in computing the eigenvalues and the eigenvectors of a matrix. Theorem 3.4 Among all linearisations y1 = W1 x,
y2 = W2 x, ˙
W1 , W2
real, non-singular
(3.12)
of the system (1.1) the one from (3.3) has minimal departure from normality and there are no other linearisations (3.12) sharing this property.
28
3 Phase Space
Proof. The linearisation (3.12) leads to the first order system y˙ = F y, F =
0 −W2 M −1 KW1−1
W1 W2−1 −1 −W2 M CW2−1
with
0 W1 L−T 1 L= 0 W2 LT2
= LAL−1
and A from (3.3). The matrices A and F have the same eigenvalues, hence −1 −T T −1 η(F )−η(A) = F 2E −A2E = η(F0 )+η(W2 L−T 2 L2 CL2 L2 W2 ), (3.13)
where A0 , F0 is the off-block diagonal part of A, F , respectively. The validity of (3.13) follows from the fact that A0 , F0 are similar and A0 is normal; the −T −1 −T T −1 same holds for L−1 and W2 L−T 2 CL2 2 L2 CL2 L2 W2 . Thus, η(F ) ≥ η(A). If the two are equal then we must have both η(F0 ) = 0
−1 −T T −1 and η(W2 L−T 2 L2 CL2 L2 W2 ) = 0.
−1 −T −1 −1 are normal. Since F0 Thus, both F0 and W2 L−T 2 L2 CL2 L2 W2 has purely imaginary spectrum, it must be skew-symmetric, whereas −1 −T −1 −1 −T W2 L−T (being similar to L−1 2 L2 CL2 L2 W2 2 CL2 ) has real nonnegative spectrum and must therefore be symmetric positive semidefinite. This is just the type of matrix A in (3.3). Q.E.D.
Taking C = 0 in (3.3) the matrix A becomes skew-symmetric and its eigenvalues are purely imaginary. They are best computed by taking the singular value decomposition (2.17). From the unitary similarity
VT 0 0 UT
LT1 L−T 2 −1 −L2 L1 0 0
V 0 0 U
=
0 Ω −Ω 0
the eigenvalues of A are seen to be ±iω1 , . . . , ±iωn . Exercise 3.5 Compute the matrix A in the damping-free case and show that eAt is unitary. Exercise 3.6 Find the phase-space matrix A for a modally damped system. Exercise 3.7 Show that different choices for L1,2 given in Remark 3.2 lead to unitarily equivalent A’s. Exercise 3.8 Show that the phase-space matrix A from (3.3) is normal, if and only if C = 0.
Chapter 4
The Singular Mass Case
If the mass matrix 𝑀 is singular no standard transformation to a first order system is possible. In fact, in this case we cannot prescribe some initial velocities and the phase-space will have dimension less than 2𝑛. This will be even more so, if the damping matrix, too, is singular; then we could not prescribe even some initial positions. In order to treat such systems we must first separate away these ‘inactive’ degrees of freedom and then arrive at phase-space matrices which have smaller dimension but their structure will be essentially the same as in the regular case studied before. Now, out of 𝑀, 𝐶, 𝐾 only 𝐾 is supposed to be positive definite while 𝑀, 𝐶 are positive semidefinite. To perform the mentioned separation it is convenient to simultaneously diagonalise the matrices 𝑀 and 𝐶 which now are allowed to be only positive semidefinite. Lemma 4.1 If 𝑀, 𝐶 are any real, symmetric, positive semidefinite matrices then there exists a real non-singular matrix 𝛷 such that 𝛷𝑇 𝑀 𝛷
and
𝛷𝑇 𝐶𝛷
are diagonal. Proof. Suppose first that 𝒩 (𝑀 ) ∩ 𝒩 (𝐶) = {0}. Then 𝑀 + 𝐶 is positive definite and there is a 𝛷 such that 𝛷𝑇 (𝑀 + 𝐶)𝛷 = 𝐼, 𝝁 diagonal. Then
𝛷𝑇 𝑀 𝛷 = 𝝁,
𝛷𝑇 𝐶𝛷 = 𝐼 − 𝝁
K. Veseli´ c, Damped Oscillations of Linear Systems, Lecture Notes in Mathematics 2023, DOI 10.1007/978-3-642-21335-9 4, © Springer-Verlag Berlin Heidelberg 2011
29
30
4 The Singular Mass Case
is diagonal as well. Otherwise, let 𝑢1 , . . . , 𝑢𝑘 be an orthonormal basis of 𝒩 (𝑀 ) ∩ 𝒩 (𝐶). Then there is an orthogonal matrix ] [ ˆ 𝑢1 ⋅ ⋅ ⋅ 𝑢 𝑘 𝑈= 𝑈 such that
] ˆ 0 𝑀 , 𝑈 𝑀𝑈 = 0 0 𝑇
[
] 𝐶ˆ 0 𝑈 𝐶𝑈 = , 0 0 𝑇
[
ˆ ) ∩ 𝒩 (𝐶) ˆ = {0} and there is a non-singular 𝛷ˆ such that 𝛷ˆ𝑇 𝑀 ˆ 𝛷ˆ where 𝒩 (𝑀 𝑇 ˆˆ ˆ and 𝛷 𝐶 𝛷 diagonal. Now set ] [ 𝛷ˆ 0 . 𝛷=𝑈 00 Q.E.D. We now start separating the ‘inactive’ variables. Using the previous lemma there is a non-singular 𝛷 such that ⎤ ⎡ ′ 𝑀1 0 0 𝑀 ′ = 𝛷𝑇 𝑀 𝛷 = ⎣ 0 0 0 ⎦ 0 00 ⎡ ′ ⎤ 𝐶1 0 0 𝐶 ′ = 𝛷𝑇 𝐶𝛷 = ⎣ 0 𝐶2′ 0 ⎦ , 0 0 0 where 𝑀1′ , 𝐶2′ are positive semidefinite (𝐶2′ may be lacking). Then set ⎡
⎤ ′ ′ ′ 𝐾12 𝐾13 𝐾11 ′𝑇 ′ ′ ⎦, 𝐾 ′ = 𝛷𝑇 𝐾𝛷 = ⎣ 𝐾12 𝐾22 𝐾23 ′𝑇 ′𝑇 ′ 𝐾13 𝐾23 𝐾33 ′ is positive definite (as a principal submatrix of the positive where 𝐾33 definite 𝐾). By setting
⎤ 𝑥′1 𝑥′ = ⎣ 𝑥′2 ⎦ , 𝑥′3 ⎡
𝑥 = 𝛷𝑥′ ,
𝐾 ′ = 𝛷𝑇 𝐾𝛷
we obtain an equivalent system ′ ′ ′ 𝑥′1 + 𝐾12 𝑥′2 + 𝐾13 𝑥′3 = 𝜑1 𝑀1′ 𝑥¨′1 + 𝐶1′ 𝑥˙ 1 + 𝐾11 ′ ′ ′ 𝐶2′ 𝑥˙ ′2 + 𝐾12𝑇 𝑥′1 + 𝐾22 𝑥′2 + 𝐾23 𝑥′3 = 𝜑2 ′ ′ 𝑇 ′ 𝑇 ′ ′ 𝐾13 𝑥1 + 𝐾23 𝑥2 + 𝐾33 𝑥′3 = 𝜑3
(4.1)
4 The Singular Mass Case
31
with 𝜑 = 𝛷𝑇 𝑓 . Eliminating 𝑥′3 from the third line of (4.1) gives ˜𝑥 ˜𝑥 ¨ 𝑀 ˜ + 𝐶˜ 𝑥 ˜˙ + 𝐾 ˜ = 𝜑(𝑡) ˜ with
[
(4.2)
[ ′ ] 𝑀1 0 ˜ 𝑀= , (4.3) 0 0 [ ] ′ ′ 𝜑1 − 𝐾13 𝐾33−1 𝜑3 𝜑˜ = , (4.4) −1 ′ 𝜑2 − 𝐾23 𝐾33 𝜑3 [ ′ ] 𝐶1 0 ˜ 𝐶= , 0 𝐶2′ ] [ ] [ ′ ′ ′−1 ′ 𝑇 −1 ′ 𝑇 ′ ′ ′ ˜ 11 𝐾 ˜ 12 𝐾 𝐾 − 𝐾 𝐾 𝐾 𝐾 − 𝐾 𝐾 𝐾 11 13 33 13 12 13 33 23 ˜ = 𝐾 ˜𝑇 𝐾 ˜ 22 = 𝐾 ′ 𝑇 − 𝐾 ′ 𝐾 ′ −1 𝐾 ′ 𝑇 𝐾 ′ − 𝐾 ′ 𝐾 ′ −1 𝐾 ′ 𝑇 . 𝐾 12 12 23 33 13 22 23 33 23 𝑥 ˜=
] 𝑥′1 , 𝑥′2
The relation
1 𝑇 1 ′ 1 ˙𝑇 ˜ ˙ 𝑥˙ 𝑀 𝑥˙ = 𝑥˙ 𝑇 𝑀 ′ 𝑥˙ ′ = 𝑥 ˜, ˜ 𝑀𝑥 2 2 2 is immediately verified. Further,
(4.5)
2 1 𝑇 1 1 ∑ ′𝑇 ′ ′ 1 ′𝑇 ′ ′ ′ ′ ′𝑇 ′ ′ 𝑥 𝐾𝑥 = 𝑥′𝑇 𝐾 ′ 𝑥′ = 𝑥𝑗 𝐾𝑗𝑘 𝑥𝑘 + 𝑥′𝑇 1 𝐾13 𝑥3 + 𝑥2 𝐾23 𝑥3 + 𝑥3 𝐾33 𝑥3 2 2 2 2 𝑗,𝑘=1
=
1 2
2 ∑
′−1 ′ ′ ′𝑇 ′ ′𝑇 ′ ′𝑇 ′ 𝑥′𝑇 𝑗 𝐾𝑗𝑘 𝑥𝑘 + 𝑥1 𝐾13 𝐾33 (𝜑3 − 𝐾13 𝑥1 − 𝐾23 𝑥2 )
𝑗,𝑘=1
′−1 ′ ′𝑇 ′ ′𝑇 ′ + 𝑥′𝑇 2 𝐾23 𝐾33 (𝜑3 − 𝐾13 𝑥1 − 𝐾23 𝑥2 )
1 ′−1 ′ ′𝑇 ′ ′𝑇 ′ ′𝑇 ′ + (𝜑𝑇3 − 𝑥′𝑇 1 𝐾13 − 𝑥2 𝐾23 )𝐾33 (𝜑3 − 𝐾13 𝑥1 − 𝐾23 𝑥3 ) 2 1 𝑇 ˜ ′−1 ′−1 ′ ′𝑇 ′ = 𝑥 ˜ 𝐾 𝑥˜ + 𝑥′𝑇 1 𝐾13 𝐾33 𝜑3 + 𝑥2 𝐾23 𝐾33 𝜑3 2 1 𝑇 ′−1 ′−1 ′𝑇 ′ + (−𝑥′𝑇 1 𝐾13 − 𝑥2 𝐾23 )𝐾33 𝜑3 + 𝜑3 𝐾33 𝜑3 2 1 𝑇 ˜ 1 ′−1 = 𝑥 ˜ 𝐾 𝑥˜ + 𝜑𝑇3 𝐾33 𝜑3 2 2 ˜ is positive definite. Exercise 4.2 Show that 𝐾 In particular, if 𝑓 = 0 (free oscillations) then 1 𝑇 1 𝑇 ˜ 𝑥 𝐾𝑥 = 𝑥 ˜ 𝐾𝑥 ˜. 2 2
(4.6)
32
4 The Singular Mass Case
Note that quite often in applications the matrices 𝑀 and 𝐶 are already diagonal and the common null space can be immediately separated off. Anyhow, we will now suppose that the system (1.1) has already the property 𝒩 (𝑀 ) ∩ 𝒩 (𝐶) = {0}. For future considerations it will be convenient to simultaneously diagonalise not 𝑀 and 𝐶 but 𝑀 and 𝐾. To this end we will use a coordinate transformation 𝛷 as close as possible to the form (2.4). Proposition 4.3 In (1.1) let the matrix 𝑀 have rank 𝑚 < 𝑛 and let 𝒩 (𝑀 )∩ 𝒩 (𝐶) = {0}. Then for any positive definite matrix 𝛺2 of order 𝑛 − 𝑚 there is a real non-singular 𝛷 such that [
] 𝐼𝑚 0 , 𝛷 𝑀𝛷 = 0 0 𝑇
[
] 𝛺12 0 𝛷 𝐾𝛷 = 𝛺 = , 0 𝛺22 𝑇
2
(4.7)
𝛺1 positive definite. Furthermore, in the matrix 𝐷 = 𝛷𝑇 𝐶𝛷 =
[
𝐷11 𝐷12 𝑇 𝐷12 𝐷22
] (4.8)
the block 𝐷22 is positive definite. Moreover, if any other 𝛷′ performs (4.7), possibly with different 𝛺1′ , 𝛺2′ then [
] 𝑈11 0 𝛷 =𝛷 , 0 𝛺2−1 𝑈22 𝛺2′ ′
where
[ 𝑈=
is unitary and
𝑈11 0 0 𝑈22
] (4.10)
𝑇 𝛺1 𝑈11 . 𝛺1′ = 𝑈11
Proof. There is certainly a Φ with ] [ 𝑀1 0 , Φ𝑇 𝑀 Φ = 0 0
(4.9)
(4.11)
Φ𝑇 𝐾Φ = 𝐼
(4.12)
where 𝑀1 is positive definite diagonal matrix of rank 𝑚; this can be done by the simultaneous diagonalisation (2.1) of the pair 𝑀, 𝐾 of symmetric matrices the second of which is positive definite and Φ is said to be 𝐾-normalised (the 𝑀 -normalisation as in (2.1) is not possible 𝑀 being singular). For given 𝛺2 −1/2 and 𝛺1 = 𝑀1 the matrix [
−1/2
𝑀1 𝛷=Φ 0
0 𝛺2
]
4 The Singular Mass Case
33
is obviously non-singular and satisfies (4.7). Moreover, the matrix 𝐷 from (4.8) satisfies ([ ]) 𝐼𝑚 0 𝒩 ∩ 𝒩 (𝐷) = {0}, 0 0 hence 𝐷22 is positive definite. Now we prove the relations (4.9)–(4.11). Suppose that we have two transformations 𝛷 and 𝛷′ both performing (4.7) with the corresponding matrices 𝛺1 , 𝛺2 , 𝛺1′ , 𝛺2′ , respectively. Then the transformations ] 𝛺1−1 0 , Φ=𝛷 0 𝛺2−1 [
′
Φ =𝛷
′
[
𝛺1′−1 0 0 𝛺2′−1
]
are 𝐾-normalised fulfilling (4.12) with 𝑀1 , 𝑀1′ , respectively. Hence, as one immediately sees, Φ′ = Φ𝑈, where
[ 𝑈=
𝑈11 0 0 𝑈22
]
is unitary and 𝑇 𝑇 𝑀1′ = 𝑈11 𝑀1 𝑈11 or, equivalently 𝛺1′−1 = 𝑈11 𝛺1−1 𝑈11 .
Hence 𝛷
′
[
𝛺1′−1 0 0 𝛺2′−1
]
(4.13)
[
] 0 𝛺1−1 𝑈11 =𝛷 . 0 𝛺2−1 𝑈22
Since the matrix 𝑈11 is orthogonal (4.13) is rewritten as 𝛺1−1 𝑈11 𝛺1′ = 𝑈11 and we have ] [ −1 0 𝛺1 𝑈11 𝛺1′ ′ , 𝛷 =𝛷 0 𝛺2−1 𝑈22 𝛺2′ [ ] 0 𝑈11 =𝛷 . 0 𝛺2−1 𝑈22 𝛺2′ Q.E.D. We now proceed to construct the phase-space formulation of (1.1) which, after the substitution 𝑥 = 𝛷𝑦, 𝛷 from (4.7), reads 𝑦¨1 + 𝐷11 𝑦˙ 1 + 𝐷12 𝑦˙ 2 + 𝛺12 𝑦1 = 𝜙1 , 𝑇 𝑦˙ 1 + 𝐷22 𝑦˙ 2 + 𝛺22 𝑦2 = 𝜙2 , 𝐷12
(4.14) (4.15)
34
4 The Singular Mass Case
with
[
𝜙1 𝜙2
]
= 𝛷𝑇 𝑓.
By introducing the new variables 𝑧1 = 𝛺1 𝑦1 ,
𝑧2 = 𝛺2 𝑦2 ,
𝑧3 = 𝑦˙1
the system (4.14) and (4.15) becomes 𝑧˙ 1 = 𝛺1 𝑧3 ( ) 𝑇 𝜙2 − 𝐷12 𝑧˙2 = 𝑧3 − 𝛺2 𝑧2 ( ) −1 𝑇 𝑧˙3 = 𝜙1 − 𝐷11 𝑧3 − 𝐷12 𝐷22 𝜙2 − 𝐷12 𝑧3 − 𝛺2 𝑧2 − 𝛺1 𝑧1 −1 𝛺2 𝐷22
(4.16) (4.17) (4.18)
or, equivalently ⎡
𝑧˙ = 𝐴𝑧 + 𝐺, where
⎤ 0 −1 ⎦, 𝐺=⎣ 𝜙2 𝛺2 𝐷22 −1 𝜙1 − 𝐷12 𝐷22 𝜙2
⎤ 0 0 𝛺1 −1 −1 𝑇 ⎦ 𝐴 = ⎣ 0 −𝛺2 𝐷22 𝛺2 −𝛺2 𝐷22 𝐷12 −1 ˆ −𝛺1 𝐷12 𝐷22 𝛺2 −𝐷 ⎡
is of order 𝑛 + 𝑚 and
(4.19)
ˆ = 𝐷11 − 𝐷12 𝐷 −1 𝐷 𝑇 𝐷 12 22
is positive semidefinite because 𝐷 itself is such. Conversely, let (4.16)–(4.18) hold and set 𝑧1 = 𝛺1 𝑦1 , 𝑧2 = 𝛺2 𝑦2 . Then by (4.16) we have 𝑦˙ 1 = 𝑧3 and (4.17) implies (4.15) whereas (4.18) and (4.15) imply (4.14). The total energy identity ∥𝑧∥2 = ∥𝑧1 ∥2 + ∥𝑧2 ∥2 + ∥𝑧3 ∥2 = 𝑥𝑇 𝐾𝑥 + 𝑥˙ 𝑇 𝑀 𝑥˙ is analogous to (3.8). As in (3.5) and (3.6) we have
and
𝑧 𝑇 𝐴𝑧 ≤ 0 for all 𝑧,
(4.20)
𝐴𝑇 = 𝐽𝐴𝐽
(4.21)
4 The Singular Mass Case
35
with
] 𝐼𝑛 0 . 𝐽= 0 −𝐼𝑚 [
(4.22)
There is a large variety of 𝛷’s performing (4.7). However, the resulting 𝐴’s in (4.19) are mutually unitarily equivalent. To prove this it is convenient to go to the inverses. For 𝐴 from (4.19) and 𝛺 = diag(𝛺1 , 𝛺2 ) the identity −1
𝐴
[
−𝛺 −1 𝐷𝛺 −1 −𝐹 = 𝐹𝑇 0𝑚
]
[
with
𝛺1−1 𝐹 = 0
] (4.23)
is immediately verified. Suppose now that 𝛷′ also performs (4.7) with the corresponding matrices 𝛺1′ , 𝛺2′ , 𝐷′ 𝐴′−1 , respectively. Then, according to (4.9) and (4.10) in Proposition 4.3, 𝐷′ = 𝛷′𝑇 𝐶𝛷′ =
[
] [ ] 𝑇 0 0 𝑈11 𝑈11 𝐷 𝑇 0 𝛺2′ 𝑈22 0 𝛺2−1 𝑈22 𝛺2′ 𝛺2−1
and, using (4.13), 𝛺
′−1
′
𝐷𝛺
′−1
] [ ] 𝑇 𝑈11 𝛺1′−1 𝛺1′−1 𝑈11 0 0 𝐷 = 𝑇 𝛺2−1 0 𝑈22 0 𝛺2−1 𝑈22 [
] [ −1 ] 𝑇 𝛺1 𝑈11 𝛺1′−1 𝑈11 0 0 𝐷 = 𝑈 𝑇 𝛺 −1 𝐷𝛺 −1 𝑈. = 𝑇 𝛺2−1 0 𝑈22 0 𝛺2−1 𝑈22 [
Similarly, 𝐹′ = Altogether, and then where the matrix
[
𝛺1′−1 0
]
[ =
𝑇 𝛺1−1 𝑈11 𝑈11 0
]
= 𝑈 −1 𝐹 𝑈11 .
ˆ ˆ −1 𝐴−1 𝑈 𝐴′−1 = 𝑈 ˆ −1 𝐴𝑈 ˆ, 𝐴′ = 𝑈 ] [ ˆ= 𝑈 0 𝑈 0 𝑈11
(4.24)
is unitary. In numerical calculations special choices of 𝛺2 might be made, for instance, the ones which possibly decrease the condition number of the transformation 𝛷. This issue could require further consideration. Exercise 4.4 Prove the property (4.20).
36
4 The Singular Mass Case
Example 4.5 Our model example from Fig. 1.1 allows for singular masses. We will consider the case with 𝑛 = 2 and 𝑚1 = 𝑚 > 0,
𝑚2 = 0,
𝑘1 , 𝑘2 > 0, 𝑘3 = 0,
𝑐𝑖 = 0,
𝑑1 = 0,
𝑑2 = 𝑑 > 0.
This gives [
] 𝑚0 𝑀= , 0 0
[
] 00 𝐶= , 0𝑑
[
] 𝑘1 + 𝑘2 −𝑘2 𝐾= . −𝑘2 𝑘2
Obviously here 𝒩 (𝑀 ) ∩ 𝒩 (𝐶) = {0}, so the construction (4.19) is possible. The columns of 𝛷 are the eigenvectors of the generalised eigenvalue problem 𝑀 𝜙 = 𝜈𝐾𝜙 whose eigenvalues are the zeros of the characteristic polynomial det(𝑀 − 𝜈𝐾) = 𝑘1 𝑘2 𝜈 2 − 𝑚𝑘2 𝜈 and they are 𝜈1 =
𝑚 , 𝑘1
𝜈2 = 0.
The corresponding eigenvectors are 𝜙1 = 𝛼1
[ ] 1 , 1
𝜙2 = 𝛼2
[ ] 0 . 1
√ We choose 𝛼1 = 𝛼2 = 1/ 𝑚. This gives 𝜔1 =
√ 𝑘1 /𝑚,
𝜔2 =
√
𝑘2 /𝑚
(we replace the symbols 𝛺1 , 𝛺2 , 𝐷𝑖𝑗 etc. from (4.19) by their lower case relateds since they are now scalars). Then [ ] [ ] 10 √ 𝛷 = 𝜙1 𝜙2 = / 𝑚, 11 [ ] [ 2 ] 10 𝜔1 0 𝑇 𝑇 𝛷 𝑀𝛷 = , 𝛷 𝐾𝛷 = . 00 0 𝜔22 Hence
[ ] 𝑑 11 𝛷 𝐶𝛷 = 𝐷 = . 𝑚 11 𝑇
4 The Singular Mass Case
37
Finally, by 𝑑ˆ = 𝑑11 − 𝑑212 /𝑑22 = 0 the phase-space matrix (4.19) reads √ ⎤ 0 0 √𝑘1 /𝑚 𝐴=⎣ −𝑘2 /𝑑 − 𝑘2 /𝑚 ⎦ . √0 √ − 𝑘1 /𝑚 𝑘2 /𝑚 0 ⎡
Exercise 4.6 If the null space of 𝑀 is known as [
] 𝑀1 0 𝑀= , 0 0
𝑀1 positive definite
try to form a phase space matrix 𝐴 in (4.19) without using the spectral decomposition (2.1) and using Cholesky decompositions instead. Exercise 4.7 Find out special properties of the matrix 𝐴 from (4.19) in the modal damping case. Exercise 4.8 Establish the relation between the matrices 𝐴 from (3.3) and from (4.19). Hint: Replace in (3.3) the matrix 𝑀 by 𝑀𝜖 = 𝑀 + 𝜖𝐼, 𝜖 > 0. build the corresponding matrix 𝐴𝜀 and find lim 𝐴−1 𝜀 .
𝜀→0
Exercise 4.9 Consider the extreme case of 𝑀 = 0 and solve the differential equation (1.1) by simultaneously diagonalising the matrices 𝐶, 𝐾. If not specified otherwise we shall by default understand the mass matrix 𝑀 as non-singular.
Chapter 5
‘Indefinite Metric’
The symmetry property of our phase-space matrices gives rise to a geometrical structure popularly called ‘an indefinite metric’ which we now will study in some detail. In doing so it will be convenient to include complex matrices as well. More precisely, from now on we will consider matrices over the field 𝛯 of real or complex numbers. Note that all considerations and formulae in Chaps. 2–4 are valid in the complex case as well. Instead of being real symmetric the matrices 𝑀, 𝐶, 𝐾 can be allowed to be Hermitian. Similarly, the vectors 𝑥, 𝑦, ... may be from 𝛯 𝑛 and real orthogonal matrices appearing there become unitary. The only change is to replace the transpose 𝑇 in 𝐴𝑇 , 𝑥𝑇 , 𝑦 𝑇 , 𝐿𝑇1 , 𝐿𝑇2 , ... by the adjoint ∗ . For a general 𝐴 ∈ 𝛯 𝑛,𝑛 the dissipativity means Re 𝑦 ∗ 𝐴𝑦 ≤ 0,
𝑦 ∈ ℂ𝑛 .
(5.1)
While taking complex Hermitian 𝑀, 𝐶, 𝐾 does not have direct physical meaning, the phase-space matrices are best studied as complex. Special cases in which only complex or only real matrices are meant will be given explicit mention. (The latter might be the case where in numerical applications one would care to keep real arithmetic for the sake of the computational efficiency.) We start from the property (3.6) or (4.21). It can be written as
where and
[𝐴𝑥, 𝑦] = [𝑥, 𝐴𝑦],
(5.2)
[𝑥, 𝑦] = 𝑦 ∗ 𝐽𝑥 = (𝐽𝑥, 𝑦)
(5.3)
𝐽 = diag(±1).
(5.4)
K. Veseli´ c, Damped Oscillations of Linear Systems, Lecture Notes in Mathematics 2023, DOI 10.1007/978-3-642-21335-9 5, © Springer-Verlag Berlin Heidelberg 2011
39
40
5 ‘Indefinite Metric’
More generally we will allow 𝐽 to be any matrix 𝐽 with the property 𝐽 = 𝐽 −1 = 𝐽 ∗ .
(5.5)
Such matrices are usually called symmetries and we will keep that terminology in our text hoping to cause not too much confusion with the term symmetry as a matrix property. In fact, without essentially altering the theory we could allow 𝐽 to be just Hermitian and non-singular. This has both advantages and shortcomings, see Chap. 11 below. The function [𝑥, 𝑦] has the properties: • [𝛼𝑥 + 𝛽𝑦, 𝑧] = 𝛼[𝑥, 𝑧] + 𝛽[𝑦, 𝑧] • [𝑥, 𝑧] = [𝑧, 𝑥] • [𝑥, 𝑦] = 0 for all 𝑦 ∈ 𝛯 𝑛 , if and only if
𝑥=0
The last property – we say [ ⋅ , ⋅ ] is non-degenerate – is a weakening of the common property [𝑥, 𝑥] > 0, whenever 𝑥 ∕= 0 satisfied by scalar products. This is why we call it an ‘indefinite scalar product’ and the geometric environment, created by it an ‘indefinite metric’. The most common form of a symmetry of order 𝑛 is [ ] 𝐼𝑚 0 𝐽= (5.6) 0 −𝐼𝑛−𝑚 but variously permuted diagonal matrices are also usual. Common nondiagonal forms are [ ] [ ] 0𝐼 0 𝑖𝐼 𝐽= , 𝐼0 −𝑖𝐼 0 or ⎡ ⎤ 𝐼00 𝐽 = ⎣0 0 𝐼 ⎦. 0𝐼0 Remark 5.1 Different forms of the symmetry 𝐽 are a matter of convenience, dictated by computational or theoretical requirements. Anyhow, one may always assume 𝐽 as diagonal with the values ±1 on its diagonal in any desired order. Indeed, there is a unitary matrix 𝑈 such that 𝐽˜ = 𝑈 ∗ 𝐽𝑈 = diag(±1) with any prescribed order of signs. If a matrix 𝐴 was, say, 𝐽-Hermitian then ˜ its ‘unitary map’ 𝐴˜ = 𝑈 ∗ 𝐴𝑈 will be 𝐽-Hermitian. The spectral properties of ˜ 𝐴 can be read-off from those for 𝐴. We call vectors 𝑥 and 𝑦 𝐽-orthogonal and write 𝑥[⊥]𝑦, if [𝑥, 𝑦] = 𝑦 ∗ 𝐽𝑥 = 0.
(5.7)
5.1 Sylvester Inertia Theorem
41
Similarly two sets of vectors 𝑆1 , 𝑆2 are called 𝐽-orthogonal – we then write 𝑆1 [⊥]𝑆2 – if (5.7) holds for any 𝑥 ∈ 𝑆1 and 𝑦 ∈ 𝑆2 . A vector is called • 𝐽-normalised, if ∣[𝑥, 𝑥]∣ = 1, • 𝐽-positive, if [𝑥, 𝑥] > 0, • 𝐽-non-negative, if [𝑥, 𝑥] ≥ 0 (and similarly for 𝐽-negative and 𝐽-nonpositive vectors), • 𝐽-definite, if it is either 𝐽-positive or 𝐽-negative, • 𝐽-neutral, if [𝑥, 𝑥] = 0. Analogous names are given to a subspace, if all of its non-zero vectors satisfy one of the conditions above. In addition, a subspace 𝒳 is called 𝐽-non-degenerate, if the only vector from 𝒳 , 𝐽-orthogonal to 𝒳 is zero. The synonyms of positive/negative/neutral type for 𝐽-positive/negative/neutral will also be used. Proposition 5.2 A subspace 𝒳 is 𝐽-definite, if and only if it is either 𝐽positive or 𝐽-negative. Proof. Let 𝑥± ∈ 𝒳 , [𝑥+ , 𝑥+ ] > 0, [𝑥− , 𝑥− ] < 0. Then for any real 𝑡 the equation 0 = [𝑥+ + 𝑡𝑥− , 𝑥+ + 𝑡𝑥− ] = [𝑥+ , 𝑥+ ] + 2𝑡 Re[𝑥+ , 𝑥− ] + 𝑡2 [𝑥− , 𝑥− ] has always a solution 𝑡 producing a 𝐽-neutral 𝑥+ + 𝑡𝑥− ∕= 0. Q.E.D. A set of vectors 𝑢1 , . . . , 𝑢𝑝 is called 𝐽-orthogonal, if none of them is 𝐽neutral and [𝑢𝑗 , 𝑢𝑘 ] = 0 for 𝑗 ∕= 𝑘 and 𝐽-orthonormal, if
∣[𝑢𝑗 , 𝑢𝑘 ]∣ = 𝛿𝑗𝑘 .
Exercise 5.3 𝐽-orthogonal vectors are linearly independent. Exercise 5.4 For any real matrix (3.5) implies (5.1).
5.1
Sylvester Inertia Theorem
Non-degenerate subspaces play an important role in the spectral theory of 𝐽-Hermitian matrices. To prepare ourselves for their study we will present some facts on Hermitian matrices which arise in the context of the so-called Sylvester inertia theorem. The simplest variant of this theorem says: If 𝐴 is Hermitian and 𝑇 non-singular then 𝑇 ∗ 𝐴𝑇 and 𝐴 have the same numbers of positive, negative and zero eigenvalues. We will here drop the non-singularity
42
5 ‘Indefinite Metric’
of 𝑇 , it will even not have to be square. At the same time, in tending to make the proof as constructive as possible, we will drop from it the notion of eigenvalue. Inertia is a notion more elementary than eigenvalues and it can be handled by rational operations only. First we prove the indefinite decomposition formula valid for any Hermitian matrix 𝐴 ∈ 𝛯 𝑛,𝑛 . It reads 𝐴 = 𝐺𝜶𝐺∗ , (5.8) where 𝐺 ∈ 𝛯 𝑛,𝑛 is non-singular and 𝜶 is diagonal. A possible construction of 𝐺, 𝜶 is obtained via the eigenvalue decomposition 𝐴 = 𝑈 𝛬𝑈 ∗ . Eigenvalue-free construction is given by Gaussian elimination without pivoting which yields 𝐴 = 𝐿𝐷𝐿∗ , where 𝐿 is lower triangular with the unit diagonal. This elimination breaks down, if in its course zero elements are encountered on the diagonal. Common row pivoting is forbidden here, because it destroys the Hermitian property. Only simultaneous permutations of both rows and columns are allowed. The process is modified to include block elimination steps. We shall describe one elimination step. The matrix 𝐴 is given as [
] 0 𝐴(𝑘−1) 𝐴= , 0 𝐴(𝑛−𝑘+1)
𝑘 = 1, . . . , 𝑛 − 1,
where 𝐴(𝑘−1) is of order 𝑘 − 1 and is void for 𝑘 = 1. We further partition 𝐴
(𝑛−𝑘+1)
] 𝐸 𝐶∗ , = 𝐶𝐵 [
where 𝐸 is square of order 𝑠 ∈ {1, 2} and is supposed to be non-singular. For 𝑠 = 1 the step is single and for 𝑠 = 2 double. Set ⎤ 𝐼𝑘−1 0 ⎦. 𝑋 = ⎣0 0 𝐼𝑠 −1 0 𝐶𝐸 𝐼𝑛−𝑘+1−𝑠 ⎡
Then
⎤ 𝐴(𝑘−1) 0 0 ⎦. 𝑋𝐴𝑋 ∗ = ⎣ 0 𝐸0 (𝑛−𝑘+1−𝑠) 0 0 𝐴 ⎡
In order to avoid clumsy indices we will describe the construction by the following algorithm (the symbol := denotes the common assigning operation).
5.1 Sylvester Inertia Theorem
43
Algorithm 5.5 𝛹 := 𝐼𝑛 ; 𝐷0 := 𝐼𝑛 ; 𝑘 := 1; while 𝑘 ≤ 𝑛 − 1 Find 𝑗 such that ∣𝑎𝑘𝑗 ∣ = max𝑖≥𝑘 ∣𝑎𝑘𝑖 ∣; If 𝑎𝑘𝑗 = 0 𝑘 := 𝑘 + 1; End if If ∣𝑎𝑘𝑘 ∣ ≥ ∣𝑎𝑘𝑗 ∣/2 > 0 Perform the single elimination step; 𝐴 := 𝑋𝐴𝑋 ∗ ; 𝛹 := 𝑋 ∗ 𝛹 ; 𝑘 := 𝑘 + 1; Else If ∣𝑎𝑗𝑗 ∣ ≥ ∣𝑎𝑘𝑗 ∣/2 > 0 Swap the 𝑘th and the 𝑗th columns and rows in 𝐴; Swap the 𝑘th and the 𝑗th columns in 𝛹 ; Perform the single elimination step; 𝐴 := 𝑋𝐴𝑋 ∗ ; 𝛹 := 𝑋 ∗ 𝛹 ; 𝑘 := 𝑘 + 1; Else Swap the 𝑘 + 1th and the 𝑗th columns and rows in 𝐴; Swap the 𝑘 + 1th and the 𝑗th columns in 𝛹 ; Perform the double elimination step; 𝐴 := 𝑋𝐴𝑋 ∗ ; 𝛹 := 𝑋 ∗ 𝛹 ; 𝑘 := 𝑘 + 2; End if End if End while The choices of steps and swappings in this algorithm secure that the necessary inversions are always possible. On exit a non-singular matrix 𝛹 is obtained with the property 𝛹 𝐴𝛹 ∗ = diag(𝐴1 , . . . , 𝐴𝑝 ), where 𝐴𝑠 is of order 𝑛𝑠 ∈ {1, 2}. In the latter case we have [
] 𝑎𝑏 𝐴𝑠 = , 𝑏𝑐
with ∣𝑏∣ ≥ 2 max{∣𝑎∣, ∣𝑐∣}.
This 2 × 2-matrix is further decomposed as follows: [ ][ ] 𝑏 1 ∣𝑏∣ 1 0 𝑌𝑠 = , 𝑎−𝑐 𝑏 − 2∣𝑏∣+𝑎−𝑐 1 1 − ∣𝑏∣
(5.9)
(5.10)
44
5 ‘Indefinite Metric’
[
then 𝑌𝑠 𝐴𝑠 𝑌𝑠∗
=
2∣𝑏∣ + 𝑎 + 𝑐 0 4∣𝑏∣2 0 − 2∣𝑏∣+𝑎−𝑐
] .
Here again all divisions are possible by virtue of the condition in (5.9).1 Finally replace 𝛹 by diag(𝑌1∗ , . . . , 𝑌𝑝∗ )𝛹 where any order-one 𝑌𝑠 equals to one. Then 𝛹 𝐴𝛹 ∗ = 𝜶, (5.11) 𝜶 as in (5.8). Now (5.8) follows with 𝐺 = 𝛹 −1 . Remark 5.6 The indefinite decomposition as we have presented it is, in fact, close to the common numerical algorithm described in [4]. The factor 1/2 appearing in line 7 of Algorithm 5.5 and further on is chosen for simplicity; in practice it is replaced by other values which increase the numerical stability and may depend on the sparsity of the matrix 𝐴. If 𝐴 is non-singular then all 𝛼𝑖 are different from zero and by replacing 𝛹 by diag(∣𝛼1 ∣−1/2 , . . . , ∣𝛼𝑛 ∣−1/2 )𝛹 in (5.11) we obtain 𝛹 𝐴𝛹 ∗ = diag(±1).
(5.12)
Theorem 5.7 (Sylvester theorem of inertia) If 𝐴 is Hermitian then the numbers of the positive, negative and zero diagonal elements of the matrix 𝜶 in (5.8) depends only on 𝐴 and not on 𝐺. Denoting these numbers by 𝜄+ (𝐴), 𝜄− (𝐴), 𝜄0 (𝐴), respectively, we have 𝜄± (𝑇 ∗ 𝐴𝑇 ) ≤ 𝜄± (𝐴) for any 𝑇 for which the above matrix product is defined. Both inequalities become equalities, if 𝑇 ∗ is injective. If 𝑇 is square and non-singular then also 𝜄0 (𝑇 ∗ 𝐴𝑇 ) = 𝜄0 (𝐴). Proof. Let
𝐵 = 𝑇 ∗ 𝐴𝑇
(5.13) ∗
be of order 𝑚. By (5.8) we have 𝐴 = 𝐺𝜶𝐺 , 𝐵 = 𝐹 𝜷𝐹 singular and 𝜶 = diag(𝜶+ , −𝜶− , 0𝑛0 ),
∗
with 𝐺, 𝐹 non-
𝜷 = diag(𝜷 + , −𝜷 − , 0𝑚0 ),
where 𝜶± , 𝜷± are positive definite diagonal matrices of order 𝑛± , 𝑚± , respectively.2 The equality (5.13) can be written as 𝑍 ∗ 𝜶𝑍 = 𝜷
(5.14)
1 One
may object that the matrix 𝑌𝑠 needs irrational operation in computing ∣𝑏∣ for a complex 𝑏, this may be avoided by replacing ∣𝑏∣ in (5.10) by, say, ∣ Re 𝑏∣ + ∣ Im 𝑏∣. 2 Note
that in (5.8) any desired ordering of the diagonals of 𝜶 may be obtained, if the columns of 𝐺 are accordingly permuted.
5.1 Sylvester Inertia Theorem
45
with 𝑍 = 𝐺∗ 𝑇 𝐹 −∗ . We partition 𝑍 as ⎡
⎤ 𝑍++ 𝑍+− 𝑍+0 𝑍 = ⎣ 𝑍−+ 𝑍−− 𝑍−0 ⎦ . 𝑍0+ 𝑍0− 𝑍00 Here 𝑍++ is of order 𝑛+ × 𝑚+ etc. according to the respective partitions in 𝜶, 𝜷. By writing (5.14) blockwise and by equating the 1, 1- and the 2, 2blocks, respectively, we obtain ∗ ∗ 𝑍++ 𝜶+ 𝑍++ − 𝑍−+ 𝜶− 𝑍−+ = 𝜷 + ,
∗ ∗ 𝑍+− 𝜶+ 𝑍+− − 𝑍−− 𝜶− 𝑍−− = −𝜷− .
∗ ∗ ∗ 𝜶+ 𝑍++ = 𝜷 + + 𝑍−+ 𝜶− 𝑍−+ and 𝑍−− 𝜶− 𝑍−− = Thus, the matrices 𝑍++ ∗ 𝜷− + 𝑍+− 𝜶+ 𝑍+− are positive definite. This is possible only if
𝑚+ ≤ 𝑛+ ,
𝑚− ≤ 𝑛− .
If 𝑇 is square and non-singular then 𝑍 is also square and non-singular. Hence applying the same reasoning to 𝑍 −∗ 𝜷𝑍 −1 = 𝜶 yields 𝑚+ = 𝑛 + ,
𝑚− = 𝑛 −
(and then, of necessity, 𝑚0 = 𝑛0 ). The proof that the numbers 𝜄+ , 𝜄− , 𝜄0 do not depend on 𝐺 in (5.8) is straightforward: in (5.13) we set 𝑇 = 𝐼 thus obtaining 𝜄± (𝐴) = 𝑛± . The only remaining case is the one with an injective 𝑇 ∗ in (5.13). Then 𝑇 𝑇 ∗ is Hermitian and positive definite and (5.13) implies 𝑇 𝐵𝑇 ∗ = 𝑇 𝑇 ∗𝐴𝑇 𝑇 ∗ , hence
𝜄± (𝐵) ≤ 𝜄± (𝐴) = 𝜄± (𝑇 𝑇 ∗𝐴𝑇 𝑇 ∗ ) = 𝜄± (𝐵).
The last assertion is obvious. Q.E.D. The triple
𝜄(𝐴) = (𝜄+ (𝐴), 𝜄− (𝐴), 𝜄0 (𝐴))
is called the inertia of 𝐴. Obviously, 𝜄+ (𝐴) + 𝜄− (𝐴) = rank(𝐴). Corollary 5.8 Let 𝐴ˆ be any principal submatrix of a Hermitian matrix 𝐴 then ˆ ≤ 𝜄± (𝐴). 𝜄± (𝐴) Corollary 5.9 If 𝐴 is block diagonal then its inertia is the sum of the inertiae of its diagonal blocks. Corollary 5.10 The inertia equals the number of the positive, negative and zero eigenvalues, respectively.
46
5 ‘Indefinite Metric’
Proof. Use the eigenvalue decomposition to obtain (5.8). Q.E.D. We turn back to the study of non-degenerate subspaces. Theorem 5.11 Let 𝒳 be a subspace of 𝛯 𝑛 and 𝑥1 , . . . , 𝑥𝑝 its basis and set ] [ 𝑋 = 𝑥1 ⋅ ⋅ ⋅ 𝑥𝑝 . Then the following are equivalent. (i) 𝒳 = ℛ(𝑋) possesses a 𝐽-orthonormal basis. (ii) 𝒳 is 𝐽-non-degenerate. (iii) The 𝐽-Gram matrix of the vectors 𝑥1 , . . . , 𝑥𝑝 𝐻 = (𝑥∗𝑖 𝐽𝑥𝑗 ) = 𝑋 ∗ 𝐽𝑋 is non-singular. (iv) The 𝐽-orthogonal companion of 𝒳 defined as 𝒳[⊥] = {𝑦 ∈ 𝛯 𝑛 : 𝑥∗ 𝐽𝑦 = 0
for all
𝑥 ∈ 𝒳}
is a direct complement of 𝒳 i.e. we have ˙ [⊥] = 𝛯 𝑛 𝒳 +𝒳 and we write
𝒳 [+]𝒳[⊥] = 𝛯 𝑛 .
In this case 𝒳[⊥] is called the 𝐽-orthogonal complement of 𝒳 . Proof. Let 𝑢1 , . . . , 𝑢𝑝 be a 𝐽-orthonormal basis in 𝒳 then 𝑥 = 𝛼1 𝑢1 + ⋅ ⋅ ⋅ + 𝛼𝑝 𝑢𝑝 is 𝐽-orthogonal to 𝒳 , if and only if 𝛼1 = ⋅ ⋅ ⋅ = 𝛼𝑝 = 0 i.e. 𝑥 = 0. Thus, (i) ⇒ (ii). To prove (ii) ⇔ (iii) note that 𝐻𝛼 = 𝑋 ∗ 𝐽𝑋𝛼 = 0 for some if and only if
𝑋 ∗ 𝐽𝑥 = 0
𝛼 ∈ 𝛯 𝑝 , 𝛼 ∕= 0
for some 𝑥 ∈ 𝒳 ,
𝑥 ∕= 0
and this means that 𝒳 is 𝐽-degenerate. Now for (ii) ⇔ (iv). The 𝐽-orthogonal companion of 𝒳 is the (standard) orthogonal complement of 𝐽𝒳 ; since 𝐽 is non-singular we have dim 𝐽𝒳 = dim 𝒳 = 𝑝 and so, dim 𝒳[⊥] = 𝑛 − 𝑝 and we have to prove 𝒳 ∩ 𝒳[⊥] = {0} but this is just the non-degeneracy of 𝒳 . Since the latter is obviously equivalent to the non-degeneracy of 𝒳[⊥] this proves (ii) ⇒ (iv). To prove
5.1 Sylvester Inertia Theorem
47
(iv) ⇔ (iii) complete the vectors 𝑥1 , . . . , 𝑥𝑝 to a basis 𝑥1 , . . . , 𝑥𝑛 of 𝛯 𝑛 such that 𝑥𝑝+1 , . . . , 𝑥𝑛 form a basis in 𝒳[⊥] and set ] [ ˜ = 𝑥1 ⋅ ⋅ ⋅ 𝑥𝑛 . 𝑋 Then the matrix
] [ ˜ =𝑋 ˜ ∗𝐽 𝑋 ˜ = 𝐻 0 𝐻 ˜ 22 0 𝐻
is non-singular. The same is then true of 𝐻 and this is (iii). It remains to prove e.g. (iii) ⇒ (i). By (5.12) 𝐻 = 𝑋 ∗ 𝐽𝑋 = 𝐹1 𝐽1 𝐹1∗ because 𝐻 is non-singular. Then the vectors 𝑢𝑗 = 𝑋𝐹1−∗ 𝑒𝑗 from 𝒳 are 𝐽 orthonormal: ∣𝑢∗𝑗 𝐽𝑢𝑘 ∣ = ∣𝑒∗𝑗 𝐽1 𝑒𝑘 ∣ = 𝛿𝑗𝑘 . Q.E.D. Corollary 5.12 Any 𝐽-orthonormal set can be completed to a 𝐽orthonormal basis in 𝛯 𝑛 . Proof. The desired basis consists of the columns of the matrix 𝑈 in the proof of Theorem 5.11. Q.E.D. Theorem 5.13 Any two 𝐽-orthonormal bases in a 𝐽-non-degenerate space have the same number of 𝐽-positive (and 𝐽-negative) vectors. Proof. Let two 𝐽-orthonormal bases be given as the columns of 𝑈 = [𝑢1 , . . . , 𝑢𝑝 ] and 𝑉 = [𝑣1 , . . . , 𝑣𝑝 ], respectively. Then
𝑈 ∗ 𝐽𝑈 = 𝐽1 , 𝐽1 = diag(±1),
𝑉 ∗ 𝐽𝑉 = 𝐽2 𝐽2 = diag(±1).
On the other hand, by ℛ(𝑈 ) = ℛ(𝑉 ) we have 𝑈 = 𝑉 𝑀, so
𝑀
non-singular,
𝐽1 = 𝑀 ∗ 𝑉 ∗ 𝐽𝑉 𝑀 = 𝑀 ∗ 𝐽2 𝑀.
By the theorem of Sylvester the non-singularity of 𝑀 implies that the number of positive and negative eigenvalues of 𝐽1 and 𝐽2 – and this is just the number of plus and minus signs on their diagonal – coincide. Q.E.D.
48
5 ‘Indefinite Metric’
Corollary 5.14 For any subspace 𝒳 ⊆ 𝛯 𝑛 we define 𝜄(𝒳 ) = (𝜄+ (𝒳 ), 𝜄0 (𝒳 ), 𝜄− (𝒳 )) = 𝜄(𝐻), where 𝐻 = 𝑋 ∗ 𝐽𝑋 and 𝑋 = [ 𝑥1 , . . . , 𝑥𝑝 ] is any basis of 𝒳 . Then 𝜄(𝒳 ) does not depend on the choice of the basis 𝑥1 , . . . , 𝑥𝑝 ∈ 𝒳 . If 𝒳 is non-degenerate then obviously 𝜄0 (𝒳 ) = 𝜄0 (𝒳[⊥] ) = 0 and 𝜄± (𝒳 ) + 𝜄± (𝒳[⊥] ) = 𝜄± (𝐽).
(5.15)
Exercise 5.15 If 𝒳1 and 𝒳2 are any two non-degenerate, mutually 𝐽orthogonal subspaces then their direct sum 𝒳 is also non-degenerate and 𝜄(𝒳 ) = 𝜄(𝒳1 ) + 𝜄(𝒳2 ) (the addition is understood elementwise).
Chapter 6
Matrices and Indefinite Scalar Products
Having introduced main notions of the geometry based on an indefinite scalar product we will now study special classes of matrices intimately connected with this scalar product. These will be the analogs of the usual Hermitian and unitary matrices. At first sight the formal analogy seems complete but the indefiniteness of the underlying scalar product often leads to big, sometimes surprising, differences. A matrix 𝐻 ∈ 𝛯 𝑛,𝑛 is called 𝐽-Hermitian (also 𝐽-symmetric, if real), if 𝐻 ∗ = JHJ
⇔
(JH)∗ = JH
or, equivalently, [𝐻𝑥, 𝑦] = 𝑦 ∗ JH𝑥 = 𝑦 ∗ 𝐻 ∗ 𝐽𝑥 = [𝑥, Hy]. It is convenient to introduce the 𝐽-adjoint 𝐴[∗] or the 𝐽-transpose 𝐴[𝑇 ] of a general matrix 𝐴, defined as 𝐴[∗] = JA∗ 𝐽,
𝐴[𝑇 ] = JA𝑇 𝐽,
respectively. In the latter case the symmetry 𝐽 is supposed to be real. Now the 𝐽-Hermitian property is characterised by 𝐴[∗] = 𝐴 or, equivalently by [𝐴𝑥, 𝑦] = [𝑥, 𝐴𝑦] for all 𝑥, 𝑦 ∈ 𝛯 𝑛 . A matrix 𝑈 ∈ 𝛯 𝑛,𝑛 is 𝐽-unitary (also 𝐽-orthogonal, if real), if 𝑈 −1 = 𝑈 [∗] = 𝐽𝑈 ∗ 𝐽
⇔
𝑈 ∗ 𝐽𝑈 = 𝐽.
K. Veseli´ c, Damped Oscillations of Linear Systems, Lecture Notes in Mathematics 2023, DOI 10.1007/978-3-642-21335-9 6, © Springer-Verlag Berlin Heidelberg 2011
49
50
6 Matrices and Indefinite Scalar Products
Obviously all 𝐽-unitaries form a multiplicative group and satisfy ∣ det 𝑈 ∣ = 1. The 𝐽-unitarity can be expressed by the identity [𝑈 𝑥, 𝑈 𝑦] = 𝑦 ∗ 𝑈 ∗ 𝐽𝑈 𝑥 = [𝑥, 𝑦]. Exercise 6.1 Prove
𝐼 [∗] = 𝐼 (𝛼𝐴 + 𝛽𝐵)[∗] = 𝛼𝐴[∗] + 𝛽𝐵 [∗] (𝐴𝐵)[∗] = 𝐵 [∗] 𝐴[∗] (𝐴[∗] )−1 = (𝐴−1 )[∗]
𝐴[∗] = 𝐴−1 ⇐⇒ A is 𝐽-unitary [ ] 𝐼 0 In the particular case 𝐽 = a 𝐽-Hermitian 𝐴 looks like 0 −𝐼 [ 𝐴= [
0𝐼 whereas for 𝐽 = 𝐼0
] 𝐴11 𝐴12 , −𝐴∗12 𝐴22
𝐴∗11 = 𝐴11 ,
𝐴∗22 = 𝐴22
]
[
the 𝐽-Hermitian is
] 𝐴11 𝐴12 𝐴= , 𝐴21 𝐴∗11
𝐴∗12 = 𝐴12 ,
𝐴∗21 = 𝐴21 .
(6.1)
By the unitary invariance of the spectral norm the condition number of a 𝐽-unitary matrix 𝑈 is 𝜅(𝑈 ) = ∥𝑈 ∥∥𝑈 −1∥ = ∥𝑈 ∥∥𝐽𝑈 ∗𝐽∥ = ∥𝑈 ∥∥𝑈 ∗ ∥ = ∥𝑈 ∥2 ≥ 1 We call a matrix jointly unitary, if it is simultaneously 𝐽-unitary and unitary. Examples of jointly unitary matrices 𝑈 are given in (3.11) and (4.24). Exercise 6.2 Prove that the following are equivalent (i) (ii) (iii) (iv) (v)
U 𝑈 𝑈 𝑈 𝑈
is is is is is
jointly unitary of order 𝑛. 𝐽-unitary and ∥𝑈 ∥ = 1. 𝐽-unitary and 𝑈 commutes with 𝐽. unitary and 𝑈 commutes with 𝐽. 𝐽-unitary and ∥𝑈 ∥2𝐸 = Tr 𝑈 ∗ 𝑈 = 𝑛.
6 Matrices and Indefinite Scalar Products
51
Example 6.3 Any matrix of the form ( ) (√ ) 𝑉1 0 𝐼 + 𝑊𝑊∗ √ 𝑊 𝑌 = 𝐻 (𝑊 ) , 𝐻 (𝑊 ) = 0 𝑉2 𝑊∗ 𝐼 + 𝑊 ∗𝑊
(6.2)
is obviously 𝐽-unitary with 𝐽 from (5.6); here 𝑊 is an 𝑚 × (𝑛 − 𝑚)-matrix and 𝑉1 , 𝑉2 are unitary. As a matter of fact, any 𝐽-unitary 𝑈 is of this form. This we will now show. Any 𝐽-unitary 𝑈 , partitioned according to (5.6) is written as [ ] 𝑈11 𝑈12 𝑈= . 𝑈21 𝑈22 By the polar 𝑈11 = 𝑊11 𝑉1 , 𝑈22 = 𝑊22 𝑉2 with 𝑊11 = √ taking √ decompositions ∗ , 𝑊 ∗ we have 𝑈11 𝑈11 = 𝑈 𝑈 22 22 22 [
𝑊11 𝑊12 𝑈= 𝑊21 𝑊22
][
] 𝑉1 0 . 0 𝑉2
Now in the product above the second factor is 𝐽-unitary (𝑉1,2 being unitary). Thus, the first factor – we call it 𝐻 – is also 𝐽-unitary, that is, 𝐻 ∗ 𝐽𝐻 = 𝐽 or, equivalently, 𝐻𝐽𝐻 ∗ = 𝐽. This is expressed as 2 ∗ 2 ∗ 𝑊11 − 𝑊21 𝑊21 = 𝐼𝑚 𝑊11 − 𝑊12 𝑊12 = 𝐼𝑚 ∗ ∗ 𝑊11 𝑊12 − 𝑊21 𝑊22 = 0 𝑊11 𝑊21 − 𝑊12 𝑊22 = 0 ∗ 2 ∗ 2 𝑊12 − 𝑊22 = −𝐼𝑛−𝑚 𝑊21 𝑊21 − 𝑊22 = −𝐼𝑛−𝑚 𝑊12
This gives (note that 𝑊11 , 𝑊22 are Hermitian positive semidefinite) √ √ ∗ ∗ 𝑊 𝑊 ∗ 𝐼𝑚 + 𝑊21 21 12 = 𝑊21 𝐼𝑛−𝑚 + 𝑊21 𝑊21 or, equivalently, ∗ ∗ ∗ 𝑊12 (𝐼𝑛−𝑚 + 𝑊12 𝑊12 )−1/2 = (𝐼𝑚 + 𝑊21 𝑊21 )−1/2 𝑊21 ∗ ∗ −1/2 ∗ ∗ = 𝑊21 (𝐼𝑛−𝑚 + 𝑊21 𝑊21 ) = 𝑊21 (𝐼𝑛−𝑚 + 𝑊12 𝑊12 )−1/2 ∗ hence 𝑊21 = 𝑊12 . Here the second equality follows from the general identity 𝐴𝑓 (𝐵𝐴) = 𝑓 (𝐴𝐵)𝐴 which we now assume as known and will address in discussing analytic matrix functions later. Now set 𝑊 = 𝑊12 and obtain (6.2).
Jointly unitary matrices are very precious whenever they can be used in computations, because their condition is equal to one. If 𝐴 is 𝐽 -Hermitian and 𝑈 is 𝐽-unitary then one immediately verifies that 𝐴′ = 𝑈 −1 𝐴𝑈 = 𝑈 [∗] 𝐴𝑈 is again 𝐽-Hermitian.
52
6 Matrices and Indefinite Scalar Products
If 𝐽 is diagonal then the columns (and also the rows) of any 𝐽-unitary matrix form a 𝐽-orthonormal basis. It might seem odd that the converse is not true. This is so because in our definition of 𝐽-orthonormality the order of the vectors plays no role. For instance the vectors ⎡ ⎤ 1 ⎣0⎦,
⎡ ⎤ 0 ⎣0⎦,
⎡ ⎤ 0 ⎣1⎦
0
1
0
are 𝐽-orthonormal with respect to ⎡
⎤
1
𝐽 = ⎣ −1 ⎦ 1 but the matrix 𝑈 built from these three vectors in that order is not 𝐽orthogonal. We rather have ⎡
⎤
1
𝑈 ∗ 𝐽𝑈 = 𝐽 ′ = ⎣ 1
⎦. −1
To overcome such difficulties we call a square matrix 𝐽, 𝐽 ′ -unitary (𝐽, 𝐽 ′ orthogonal, if real), if 𝑈 ∗ 𝐽𝑈 = 𝐽 ′ , (6.3) where 𝐽 and 𝐽 ′ are symmetries. If in (6.3) the orders of 𝐽 ′ and 𝐽 do not coincide we call 𝑈 a 𝐽, 𝐽 ′ -isometry. If both 𝐽 and 𝐽 ′ are unit matrices this is just a standard isometry. Exercise 6.4 Any 𝐽, 𝐽 ′ - isometry is injective. Proposition 6.5 If 𝐽, 𝐽 ′ are symmetries and 𝑈 is 𝐽, 𝐽 ′ -unitary then 𝐽 and 𝐽 ′ are unitarily similar: 𝐽 ′ = 𝑉 −1 𝐽𝑉, and
𝑉
𝑈 = 𝑊 𝑉,
unitary
(6.4) (6.5)
where 𝑊 is 𝐽-unitary. Proof. The eigenvalues of both 𝐽 and 𝐽 ′ consists of ± ones. By (6.3) 𝑈 must be non-singular but then (6.3) and the theorem of Sylvester imply that 𝐽 and 𝐽 ′ have the same eigenvalues including multiplicities, so they are unitarily similar. Now (6.5) follows from (6.4). Q.E.D.
6 Matrices and Indefinite Scalar Products
53
If 𝐴 is 𝐽-Hermitian and 𝑈 is 𝐽, 𝐽 ′ -unitary then 𝐴′ = 𝑈 −1 𝐴𝑈 is 𝐽 ′ -Hermitian. Indeed, 𝐴′∗ = 𝑈 ∗ 𝐴∗ 𝑈 −∗ = 𝑈 ∗ AJUJ′ . Here 𝐴𝐽, and therefore 𝑈 ∗ 𝐴𝐽𝑈 is Hermitian so 𝐴′∗ is 𝐽 ′ -Hermitian. Using only 𝐽-unitary similarities the 𝐽-Hermitian matrix ⎡
⎤ 1 0 −5 𝐴 = ⎣ 0 2 0⎦, −5 0 1
⎡
⎤
1
𝐽 =⎣ 1
⎦
(6.6)
−1
cannot be further simplified, but using the 𝐽, 𝐽 ′ -unitary matrix ⎡
⎤ 100 𝛱 = ⎣0 0 1⎦, 010
⎡
1
⎤
𝐽 ′ = ⎣ −1 ⎦ 1
we obtain the more convenient block-diagonal form ⎡
⎤ 1 −5 0 𝛱 −1 𝐴𝛱 = ⎣ −5 1 0 ⎦ 0 02 which may have computational advantages. Since any 𝐽, 𝐽 ′ -unitary matrix 𝑈 is a product of a unitary matrix and a 𝐽-unitary matrix we have 𝜅(𝑈 ) = ∥𝑈 ∥2 ≥ 1, where the equality is attained, if and only if 𝑈 is unitary. Mapping by a 𝐽, 𝐽 ′ -unitary matrix 𝑈 preserves the corresponding ‘indefinite geometries’ e.g. • If 𝑥′ = 𝑈 𝑥, 𝑦 ′ = 𝑈 𝑦 then 𝑥∗ 𝐽𝑦 = 𝑥′∗ 𝐽 ′ 𝑦 ′ , 𝑥∗ 𝐽𝑥 = 𝑥′∗ 𝐽 ′ 𝑥′ • If 𝒳 is a subspace and 𝒳 ′ = 𝑈 𝒳 then 𝒳 ′ is 𝐽 ′ -non-degenerate, if and only if 𝒳 is 𝐽-non-degenerate (the same with ‘positive’, ‘non-negative’, ‘neutral’ etc.) • The inertia does not change: 𝜄± (𝒳 ) = 𝜄′± (𝒳 ′ ), where 𝜄′± is related to 𝐽 ′ .
54
6 Matrices and Indefinite Scalar Products
The set of 𝐽, 𝐽 ′ -unitary matrices is not essentially larger than the one of standard 𝐽-unitaries but it is often more convenient in numerical computations. Exercise 6.6 Find all real 𝐽-orthogonals and all complex 𝐽-unitaries of order 2 with [ ] [ ] 1 0 01 𝐽= or 𝐽 = . 0 −1 10 Exercise 6.7 A matrix is called jointly Hermitian, if it is both Hermitian and 𝐽-Hermitian. Prove that the following are equivalent (i) H is jointly Hermitian. (ii) 𝐻 is 𝐽-Hermitian and it commutes with 𝐽. (iii) 𝐻 is Hermitian and it commutes with 𝐽.
Chapter 7
Oblique Projections
In the course of these lectures we will often have to deal with non-orthogonal (oblique) projections, so we will collect here some useful facts about them. Any square matrix of order 𝑛 is called a projection (or projector), if 𝑃 2 = 𝑃. We list some obvious properties of projections. Proposition 7.1 If 𝑃 is a projection then the following hold. (i) ℛ(𝑃 ) = {𝑥 ∈ 𝛯 𝑛 : 𝑃 𝑥 = 𝑥}. (ii) 𝑄 = 𝐼 − 𝑃 is also a projection and 𝑄 + 𝑃 = 𝐼,
𝑃 𝑄 = 𝑄𝑃 = 0,
˙ (iii) ℛ(𝑃 )+ℛ(𝑄) = 𝛯 𝑛 , ℛ(𝑃 ) = 𝒩 (𝑄), ℛ(𝑄) = 𝒩 (𝑃 ). (iv) If 𝑥1 ,[. . . , 𝑥𝑝 is a] basis in ℛ(𝑃 ) and 𝑥𝑝+1 , . . . , 𝑥𝑛 a basis in ℛ(𝑄) then 𝑋 = 𝑥1 ⋅ ⋅ ⋅ 𝑥𝑛 is non-singular and [ ] 𝐼𝑝 0 −1 𝑋 𝑃𝑋 = . 0 0 (v) rank(𝑃 ) = Tr(𝑃 ). (vi) A set of projections 𝑃1 , . . . , 𝑃𝑞 is called a decomposition of the identity if 𝑃1 + ⋅ ⋅ ⋅ + 𝑃𝑞 = 𝐼,
𝑃𝑖 𝑃𝑗 = 𝑃𝑗 𝑃𝑖 = 𝑃𝑗 𝛿𝑖𝑗 .
To any such decomposition there corresponds the direct sum ˙ ⋅ ⋅ ⋅ +𝒳 ˙ 𝑞 = 𝛯𝑛 𝒳1 + with 𝒳𝑖 = ℛ(𝑃𝑖 ) and vice versa. (vii) 𝑃 ∗ is also a projection and ℛ(𝑃 ∗ ) = 𝒩 (𝑄)⊥ . The proofs are straightforward and are omitted. K. Veseli´ c, Damped Oscillations of Linear Systems, Lecture Notes in Mathematics 2023, DOI 10.1007/978-3-642-21335-9 7, © Springer-Verlag Berlin Heidelberg 2011
55
56
7 Oblique Projections
Exercise 7.2 If 𝑃1 , 𝑃2 are projections then 1. ℛ(𝑃1 ) ⊆ ℛ(𝑃2 ) is equivalent to 𝑃1 = 𝑃2 𝑃1 ;
(7.1)
2. ℛ(𝑃1 ) = ℛ(𝑃2 ) is equivalent to 𝑃1 = 𝑃2 𝑃1
&
𝑃2 = 𝑃1 𝑃2 ;
(7.2)
3. If, in addition, both 𝑃1 , 𝑃2 are Hermitian or 𝐽-Hermitian then (7.2) implies 𝑃2 = 𝑃1 . Exercise 7.3 Any two decompositions of the identity 𝑃1 , . . . , 𝑃𝑞 , 𝑃1′ , . . . , 𝑃𝑞′ for which Tr 𝑃𝑗 = Tr 𝑃𝑗′ , 𝑗 = 1, . . . , 𝑞, are similar, that is, there exists a non-singular 𝑆 with 𝑆 −1 𝑃𝑖 𝑆 = 𝑃𝑖′ ,
𝑖 = 1, . . . , 𝑞
if all 𝑃𝑖 , 𝑃𝑖′ are orthogonal projections then 𝑆 can be chosen as unitary. as
Define 𝜃 ∈ (0, 𝜋/2] as the minimal angle between the subspaces 𝒳 and 𝒳 ′ 𝜃 = min{arccos ∣𝑥∗ 𝑦∣, 𝑥 ∈ 𝒳 , 𝑦 ∈ 𝒳 ′ , ∥𝑥∥ = ∥𝑦∥ = 1} = arccos(max{∣𝑥∗ 𝑦∣, 𝑥 ∈ 𝒳 , 𝑦 ∈ 𝒳 ′ , ∥𝑥∥ = ∥𝑦∥ = 1}).
(7.3) (7.4)
A simpler formula for the angle 𝜃 is cos 𝜃 = ∥ℙℙ′ ∥,
(7.5)
where ℙ, ℙ′ are the orthogonal projections onto 𝒳 , 𝒳 ′ , respectively. Indeed, ∣(ℙ𝑥)∗ ℙ′ 𝑦∣ 𝑥,𝑦∕=0 ∥𝑥∥∥𝑦∥
∥ℙℙ′ ∥ = max
and this maximum is taken on a pair 𝑥 ∈ 𝒳 , 𝑦 ∈ 𝒳 ′ . In fact, for a general 𝑥 we have ∥𝑃 𝑥∥ ≤ ∥𝑥∥ and ∣(ℙ𝑥)∗ ℙ′ 𝑦∣ ∣(ℙ𝑥)∗ ℙ′ 𝑦∣ ∣(ℙ(ℙ𝑥))∗ ℙ′ 𝑦∣ ≤ = ∥𝑥∥∥𝑦∥ ∥ℙ𝑥∥∥𝑦∥ ∥ℙ𝑥∥∥𝑦∥ and similarly with 𝑦. Thus,
7 Oblique Projections
57
{
′
∥ℙℙ ∥ = max { = max
∣(ℙ𝑥)∗ ℙ′ 𝑦∣ , ∥ℙ𝑥∥∥ℙ′ 𝑦∥ ∣𝑢∗ 𝑣∣ , ∥𝑢∥∥𝑣∥
= max{∣𝑢∗ 𝑣∣,
}
′
ℙ𝑥, ℙ 𝑦 ∕= 0 𝑣 ∈ 𝒳 ′,
𝑢 ∈ 𝒳,
𝑢 ∈ 𝒳,
(7.6)
𝑣 ∈ 𝒳 ′,
} 𝑢, 𝑣 ∕= 0
(7.7)
∥𝑢∥ = ∥𝑣∥ = 1}.
(7.8)
which proves (7.5), Theorem 7.4 Let 𝑃 be a projection and let ℙ be the orthogonal projection onto ℛ(𝑃 ). Then for 𝑄 = 𝐼 − 𝑃 and ℚ = 𝐼 − ℙ we have ∥𝑃 ∥ = ∥𝑄∥ =
1 , sin 𝜃
∥𝑃 ∗ − 𝑃 ∥ = ∥𝑄∗ − 𝑄∥ = ∥ℙ − 𝑃 ∥ = ∥ℚ − 𝑄∥ = cot 𝜃,
(7.9) (7.10)
where 𝜃 is the minimal angle between ℛ(𝑃 ) and ℛ(𝑄). Proof. Take any orthonormal basis of ℛ(𝑃 ) and complete it to an orthonormal basis of 𝛯 𝑛 . These vectors are the columns of a unitary matrix 𝑈 for which [ ] 𝐼0 ℙ′ = 𝑈 ∗ ℙ𝑈 = , 00 [ ] 𝑃11 𝑃12 ′ ∗ 𝑃 = 𝑈 𝑃𝑈 = , 𝑃21 𝑃22 where 𝑃 ′ is again a projection with ℛ(𝑃 ′ ) = ℛ(ℙ′ ). Now apply (7.2). ℙ′ 𝑃 ′ = 𝑃 ′ gives [
𝑃11 𝑃12 0 0
]
[
𝑃11 𝑃12 = 𝑃21 𝑃22
] ⇒
𝑃21 = 𝑃22 = 0,
⇒
𝑃11 = 𝐼.
whereas 𝑃 ′ ℙ′ = ℙ′ gives [
] [ ] 𝑃11 0 𝐼0 = 0 0 00
Thus, we can assume that 𝑃 and ℙ already are given as [
] [ ] ] 𝐼𝑋 𝐼 [ 𝑃 = = 𝐼𝑋 0 0 0 [ ] 𝐼0 ℙ= . 00
(7.11) (7.12)
58
7 Oblique Projections
Furthermore,
[ 𝑄=𝐼 −𝑃 =
0 −𝑋 0 𝐼
]
[ =
] ] −𝑋 [ 0𝐼 𝐼
whereas the orthogonal projection onto 𝒩 (𝑃 ) = ℛ(𝑄) is [ ] [ ] ˆ = −𝑋 (𝐼 + 𝑋 ∗ 𝑋)−1 −𝑋 ∗ 𝐼 . ℙ 𝐼
(7.13)
According to (7.5) we have ([
ˆ = spr cos 𝜃 = ∥ℙℙℙ∥ 2
] ) [ ] −𝑋 ∗ −1 ∗ (𝐼 + 𝑋 𝑋) −𝑋 0 0
= spr((𝐼 + 𝑋 ∗ 𝑋)−1 𝑋 ∗ 𝑋). To evaluate the last expression we use the singular value decomposition 𝑋 = 𝑈 𝜉𝑉 ∗
(7.14)
with 𝑈, 𝑉 unitary and 𝜉 diagonal (but not necessarily square). Then 𝑋 ∗ 𝑋 = 𝑉 𝜉 ∗ 𝜉𝑉 ∗ , 𝜉 ∗ 𝜉 = diag(∣𝜉1 ∣2 , ∣𝜉2 ∣2 , . . .) and ∣𝜉𝑖 ∣2 max𝑖 ∣𝜉𝑖 ∣2 = 1 + ∣𝜉𝑖 ∣2 1 + max𝑖 ∣𝜉𝑖 ∣2
cos2 𝜃 = max 𝑖
=
∥𝑋∥2 1 =1− 1 + ∥𝑋∥2 1 + ∥𝑋∥2
Hence
∥𝑋∥ = cot 𝜃.
Now ([
] [ ] ) ] 𝐼 [ ] 𝐼 [ ∥𝑃 ∥ = spr 𝐼0 𝐼𝑋 𝑋∗ 0 ([ ]) 𝐼 + 𝑋𝑋 ∗ 0 = spr = 1 + ∥𝑋∥2 0 0 2
=
1 sin2 𝜃
Then obviously ∥𝑃 ∥ = ∥𝑄∥ and (7.9) holds.
(7.15)
7 Oblique Projections
59
Also
[ ]
0 𝑋
= ∥𝑋∥ = cot 𝜃, ∥𝑃 − 𝑃 ∥ =
𝑋∗ 0
[ ]
0𝑋
∥ℙ − 𝑃 ∥ =
0 0 = cot 𝜃 ∗
and similarly for ∥𝑄∗ − 𝑄∥ and ∥ℚ − 𝑄∥. Q.E.D. Corollary 7.5
∥𝑃 ∥ = 1
⇔
𝑃∗ = 𝑃
i.e. a projection is orthogonal, if and only if its norm equals one. Theorem 7.6 If 𝑃 is a projection and ℙ the orthogonal projection onto ℛ(𝑃 ) then there exists a non-singular 𝑆 such that 𝑆 −1 𝑃 𝑆 = ℙ and √ 1 (2 + cot2 𝜃 + 4 cot2 𝜃 + cot4 𝜃) 2 √ √ 1 = (2 + ∥𝑃 ∥2 − 1 + ∥𝑃 ∥2 − 1 3 + ∥𝑃 ∥2 ). 2
∥𝑆∥∥𝑆 −1∥ =
(7.16)
Proof. As in the proof of Theorem 7.4 we may assume that 𝑃, ℙ already have the form (7.11) and (7.12), respectively. Set [
] 𝐼 −𝑋 𝑆= . 0 𝐼 Then ][ ] 𝐼𝑋 𝐼 −𝑋 0 0 0 𝐼 [ ][ ] [ ] 𝐼𝑋 𝐼0 𝐼0 = = = ℙ. 0 𝐼 00 00
𝑆 −1 𝑃 𝑆 =
[
𝐼𝑋 0 𝐼
][
We compute the condition number of 𝑆: ] 𝐼 −𝑋 𝑆 𝑆= . −𝑋 ∗ 𝐼 + 𝑋 ∗ 𝑋 ∗
[
60
7 Oblique Projections
Then using (7.14) [
𝑈 0 𝑆 𝑆= 0 𝑉 ∗
][
𝐼 −𝜉 −𝜉 𝐼 + 𝜉 ∗ 𝜉
][
] 𝑈 0 , 0 𝑉∗
so those eigenvalues of 𝑆 ∗ 𝑆 which are different from 1 are given by 𝜆+ 𝑖
=
𝜆− 𝑖 =
2 + 𝜉𝑖2 +
√ 2
4𝜉𝑖2 + 𝜉𝑖4
> 1,
1 < 1, 𝜆+ 𝑖
where 𝜉𝑖 are the diagonal elements of 𝜉. Hence ∗
𝜆𝑚𝑎𝑥 (𝑆 𝑆) = 𝜆𝑚𝑖𝑛 (𝑆 ∗ 𝑆) =
2 + ∥𝑋∥2 +
√
4∥𝑋∥2 + ∥𝑋∥4 , 2
1 𝜆𝑚𝑎𝑥 (𝑆 ∗ 𝑆)
and √
𝜆𝑚𝑎𝑥 (𝑆 ∗ 𝑆) 𝜆𝑚𝑖𝑛 (𝑆 ∗ 𝑆) √ 2 + cot2 𝜃 + 4 cot2 𝜃 + cot4 𝜃 = 2
𝜅(𝑆) =
and this proves (ii). Q.E.D. This theorem automatically applies to any decomposition of the identity consisting of two projections 𝑃 and 𝐼 − 𝑃 . It can be recursively applied to any decomposition of the identity 𝑃1 , . . . , 𝑃𝑞 but the obtained estimates get more and more clumsy with the growing number of projections, and the condition number of the transforming matrix 𝑆 will also depend on the space dimension 𝑛. Exercise 7.7 Let 𝑃 be a projection and 𝐻 = 𝐼 − 2𝑃 . Prove (i) 𝐻 2 = 𝐼, √ (ii) 𝜅(𝐻) = 1 + 2(∥𝑃 ∥2 − 1) + 2∥𝑃 ∥ ∥𝑃 ∥2 − 1. Hint: Use the representation for 𝑃 from Theorem 7.4.
Chapter 8
𝑱 -Orthogonal Projections
𝐽-orthogonal projections are intimately related to the 𝐽-symmetry of the phase space matrices which govern damped systems. In this chapter we study these projections in some detail. A projection 𝑃 is called 𝐽-orthogonal if the matrix 𝑃 is 𝐽-Hermitian i.e. if 𝑃 [∗] = 𝐽𝑃 ∗ 𝐽 = 𝑃. A subspace 𝒳 is said to possess a 𝐽-orthogonal projection 𝑃 , if 𝒳 = ℛ(𝑃 ). Theorem 8.1 (i) A subspace 𝒳 ⊆ 𝛯 𝑛 is 𝐽-non-degenerate, if and only if it possesses a 𝐽-orthogonal projection 𝑃 . (ii) Any 𝐽-orthogonal projection 𝑃 can be represented as 𝑃 = 𝑈 𝐽 ′ 𝑈 ∗ 𝐽, (8.1) where 𝑈 is a 𝐽, 𝐽 ′ -isometry of type 𝑛 × 𝑚 and ℛ(𝑃 ) = ℛ(𝑈 ). (iii) The 𝐽-orthogonal projection 𝑃 is uniquely determined by 𝒳 . (iv) We have 𝜄± (𝐽𝑃 ) = 𝜄± (𝐽 ′ ) = 𝜄± (𝒳 ).
(8.2)
(v) 𝑄 = 𝐼 − 𝑃 is the 𝐽-orthogonal projection onto 𝒳[⊥] . Proof. (i) Let 𝑃 be a 𝐽-orthogonal projection and 𝒳 = ℛ(𝑃 ). Then for 𝑥 ∈ 𝒳 and 𝑧 ∈ 𝛯 𝑛 [𝑥, 𝑃 𝑧] = [𝑃 𝑥, 𝑧] = [𝑥, 𝑧] holds. Thus, 𝑥[⊥]𝒳 is equivalent to 𝑥[⊥]𝛯 𝑛 . Since 𝛯 𝑛 is non-degenerate the same is true of 𝒳 . Conversely, let 𝒳 be non-degenerate of dimension 𝑝 < 𝑛. Then by Theorem 5.11 and Corollary 5.12 there is a 𝐽-orthonormal basis 𝑢1 , . . . , 𝑢𝑛 of 𝛯 𝑛 such that 𝑢1 , . . . , 𝑢𝑝 spans 𝒳 . Then 𝑃 , defined as { 𝑃 𝑢𝑗 =
𝑢𝑗 , 𝑗 ≤ 𝑝 0, 𝑗 > 𝑝
is the wanted projection. This proves (i). K. Veseli´ c, Damped Oscillations of Linear Systems, Lecture Notes in Mathematics 2023, DOI 10.1007/978-3-642-21335-9 8, © Springer-Verlag Berlin Heidelberg 2011
61
62
8 𝐽-Orthogonal Projections
(ii) If 𝑃 is given by (8.1), (6.3) then 𝑃 2 = 𝑈 𝐽 ′ 𝑈 ∗ 𝐽𝑈 𝐽 ′ 𝑈 ∗ 𝐽 = 𝑈 𝐽 ′ 𝐽 ′ 𝐽 ′ 𝑈 ∗ 𝐽 = 𝑃, 𝑃 [∗] = 𝐽𝑃 ∗ 𝐽 = 𝐽 2 𝑈 𝐽 ′ 𝑈 ∗ 𝐽 = 𝑈 𝐽 ′ 𝑈 ∗ 𝐽 = 𝑃 hence 𝑃 is a 𝐽-orthogonal projection. The injectivity of 𝑈 follows from (6.3), hence also rank(𝑈 ) = 𝑝 := rank(𝑃 ). Conversely, let 𝑃 be a 𝐽orthogonal projection. Then, according to Theorem 5.11 𝒳 = ℛ(𝑃 ) is 𝐽-non-degenerate and there is a 𝐽-orthonormal basis 𝑢1 , . . . , 𝑢𝑛 of 𝛯 𝑛 such that the first 𝑝 vectors span 𝒳 . With 𝑈 = [ 𝑢1 , . . . , 𝑢𝑝 ] we have 𝑈 ∗ 𝐽𝑈 = 𝐽 ′ = diag(±1).
(8.3)
The wanted projection is 𝑃 = 𝑈 𝐽 ′𝑈 ∗𝐽
(8.4)
Obviously 𝒳 ⊆ ℛ(𝑈 ). This inclusion is, in fact equality because rank(𝑈 ) = 𝑝. This proves (ii). (iii) If there are two 𝐽-orthogonal projections with the same range 𝑈 𝐽 ′ 𝑈 ∗ 𝐽,
𝑉 𝐽2 𝑉 ∗ 𝐽,
ℛ(𝑉 ) = ℛ(𝑈 )
then 𝑉 = 𝑈 𝑀 for some non-singular 𝑀 and from 𝐽2 = 𝑉 ∗ 𝐽𝑉 = 𝑀 ∗ 𝑈 ∗ 𝐽𝑈 𝑀 = 𝑀 ∗ 𝐽 ′ 𝑀 it follows
𝑉 𝐽2 𝑉 ∗ 𝐽 = 𝑈 𝑀 𝐽2 𝑀 ∗ 𝑈 ∗ 𝐽 = 𝑈 𝐽 ′ 𝑈 ∗ 𝐽.
This proves (iii). (iv) Let 𝑃 = 𝑈 𝐽 ′ 𝑈 ∗ 𝐽. The second equality in (8.2) follows, if in Corollary 5.14 we take 𝑋 as 𝑈 . Further, since 𝐽 is a symmetry, 𝜄± (𝐽𝑃 ) = 𝜄± (𝑈 𝐽 ′ 𝑈 ∗ ) and since 𝑈 is injective, by the Sylvester inertia theorem, 𝜄± (𝑈 𝐽 ′ 𝑈 ∗ ) = 𝜄± (𝐽 ′ ). This proves (iv). The assertion (v) is obvious. Q.E.D.
8 𝐽-Orthogonal Projections
63
Let 𝑃1 , . . . , 𝑃𝑟 be 𝐽-orthogonal projections with the property 𝑃𝑖 𝑃𝑗 = 𝛿𝑖𝑗 𝐼 then their sum 𝑃 is again a 𝐽-orthogonal projection and we say that the system 𝑃1 , . . . , 𝑃𝑟 is a 𝐽-orthogonal decomposition of 𝑃 . If 𝑃 = 𝐼 then we speak of a 𝐽-orthogonal decomposition of the identity. To any 𝐽-orthogonal decomposition there corresponds a 𝐽-orthogonal sum of their ranges ℛ(𝑃 ) = ℛ(𝑃1 )[+] ⋅ ⋅ ⋅ [+]ℛ(𝑃𝑟 ) and vice versa. The proof is straightforward and is left to the reader. Theorem 8.2 Let 𝑃1 , . . . , 𝑃𝑞 be a 𝐽-orthogonal decomposition of the identity and 𝑛𝑗 = Tr(𝑃𝑗 ) the dimensions of the respective subspaces. Then there exists 𝐽, 𝐽 0 -unitary 𝑈 such that ⎡ ⎢ 𝐽0 = ⎣
𝐽10
⎤ ..
⎥ ⎦,
.
𝜄± (𝐽𝑗0 ) = 𝜄± (𝐽𝑃𝑗 )
(8.5)
𝐽𝑝0 and
⎡ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ 𝑈 −1 𝑃𝑗 𝑈 = 𝑃𝑗0 = ⎢ ⎢ ⎢ ⎢ ⎢ ⎣
0
⎤ ..
⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥. ⎥ ⎥ ⎥ ⎥ ⎦
. 0 𝐼𝑛𝑗
0
..
. 0
Proof. By Theorem 8.1 we may write 𝑃𝑗 = 𝑈𝑗 𝐽𝑗0 𝑈𝑗∗ 𝐽,
𝑈𝑗∗ 𝐽𝑈𝑗 = 𝐽𝑗0 .
[ ] Then 𝑈 = 𝑈1 ⋅ ⋅ ⋅ 𝑈𝑝 obviously satisfies ⎡
⎡ 0 ⎤ ⎤ 𝑈1∗ 𝐽1 ] [ ⎢ ⎢ ⎥ ⎥ 0 .. 𝑈 ∗ 𝐽𝑈 = ⎣ ... ⎦ 𝐽 𝑈1 ⋅ ⋅ ⋅ 𝑈𝑝 = ⎣ ⎦ =: 𝐽 . 𝐽𝑝0 𝑈𝑝∗ and
(8.6)
64
8 𝐽-Orthogonal Projections
𝑈 −1 𝑃𝑗 𝑈 = 𝐽 ′ 𝑈 ∗ 𝐽𝑈𝑗 𝐽𝑗0 𝑈𝑗∗ 𝐽𝑈 ⎡ 0 ∗⎤ 𝐽1 𝑈1 [ ] ⎢ .. ⎥ = ⎣ . ⎦ 𝐽𝑈𝑗 𝐽𝑗0 𝑈𝑗∗ 𝐽𝑈1 ⋅ ⋅ ⋅ 𝐽𝑈𝑝
𝐽𝑝0 𝑈𝑝∗ ⎡ ⎤ 0 ⎢ . ⎥ ⎢ .. ⎥ ⎢ ⎥ ⎢ 0 ⎥ ⎢ ⎥[ ] ⎢ ⎥ = ⎢ 𝐽𝑗0 ⎥ 0 ⋅ ⋅ ⋅ 0 𝐽𝑗0 0 ⋅ ⋅ ⋅ 0 = 𝑃𝑗0 . ⎢ ⎥ ⎢ 0 ⎥ ⎢ ⎥ ⎢ .. ⎥ ⎣ . ⎦ 0
Q.E.D. Corollary 8.3 Any two 𝐽-orthogonal decompositions of the identity 𝑃1 , . . . , 𝑃 𝑞
and
𝑃1′ , . . . , 𝑃𝑞′
satisfying 𝜄± (𝐽𝑃𝑗 ) = 𝜄± (𝐽𝑃𝑗′ ) are 𝐽-unitarily similar: 𝑃𝑗′ = 𝑈 −1 𝑃𝑗 𝑈,
𝑈 𝐽 -unitary.
Theorem 8.4 Let 𝑃, 𝑃 ′ be 𝐽-orthogonal projections and ∥𝑃 ′ −𝑃 ∥ < 1. Then there is a 𝐽-unitary 𝑈 such that 𝑃 ′ = 𝑈 −1 𝑃 𝑈.
(8.7)
Proof. The matrix square root [ ]−1/2 𝑍 = 𝐼 − (𝑃 ′ − 𝑃 )2 is defined by the known binomial series 𝑍=
∞ ∑ 𝑘=0
(−1)𝑘
( ) 𝛼 (𝑃 ′ − 𝑃 )2𝑘 , 𝑘
( ) 𝛼 𝛼(𝛼 − 1) ⋅ ⋅ ⋅ (𝛼 − 𝑘 + 1) = , 𝑘 𝑘!
which converges because of ∥𝑃 ′ − 𝑃 ∥ < 1. Moreover, 𝑍 is 𝐽-Hermitian and (𝑃 ′ − 𝑃 )2 (and therefore 𝑍) commutes with both 𝑃 and 𝑃 ′ . Set 𝑈 = 𝑍 [𝑃 𝑃 ′ + (𝐼 − 𝑃 )(𝐼 − 𝑃 ′ )] ,
8 𝐽-Orthogonal Projections
then
65
𝑃 𝑈 = 𝑍𝑃 𝑃 ′ = 𝑍 [𝑃 𝑃 ′ + (𝐼 − 𝑃 )(𝐼 − 𝑃 ′ )] 𝑃 ′ = 𝑈 𝑃 ′ ,
so (8.7) holds. To prove 𝐽-unitarity we compute 𝑈 ∗ 𝐽𝑈 = 𝐽𝑍 [𝑃 ′ 𝑃 + (𝐼 − 𝑃 ′ )(𝐼 − 𝑃 )] [𝑃 𝑃 ′ + (𝐼 − 𝑃 )(𝐼 − 𝑃 ′ )] 𝐽𝑍 [ ] = 𝐽𝑍 𝐼 − (𝑃 ′ − 𝑃 )2 𝑍 = 𝐽. Q.E.D. Corollary 8.5 Under the conditions of the theorem above we have 𝜄± (𝐽𝑃 ′ ) = 𝜄± (𝐽𝑃 ). In particular, if 𝐽𝑃 is positive or negative semidefinite then 𝐽𝑃 ′ will be the same. Also, a continuous 𝐽-orthogonal-projection valued function 𝑃 (𝑡) for 𝑡 from a closed real interval, cannot change the inertia of 𝐽𝑃 (𝑡). Proof. Apply the Sylvester inertia theorem to 𝐽𝑃 ′ = 𝐽𝑈 −1 𝑃 𝑈 = 𝑈 ∗ 𝐽𝑃 𝑈. Then cover the interval by a finite number of open intervals such that within each of them ∥𝑃 (𝑡) − 𝑃 (𝑡′ )∥ < 1 holds and apply Theorem 8.4. Q.E.D. Exercise 8.6 Let 𝑃1 , 𝑃2 be projections such that 𝑃1 is 𝐽-orthogonal and 𝑃2 𝐽 ′ − 𝑜𝑟𝑡ℎ𝑜𝑔𝑜𝑛𝑎𝑙. Then they are 𝐽, 𝐽 ′ -unitarily similar, if and only if 𝜄± (𝐽 ′ 𝑃2 ) = 𝜄± (𝐽𝑃1 ). Exercise 8.7 A non-degenerate subspace is definite, if and only if it possesses a 𝐽-orthogonal projection 𝑃 with one of the properties 𝑥∗ 𝐽𝑃 𝑥 ≥ 0 or 𝑥∗ 𝐽𝑃 𝑥 ≤ 0 for all 𝑥 ∈ 𝛯 𝑛 that is, the matrix 𝐽𝑃 or −𝐽𝑃 is positive semidefinite. A 𝐽-orthogonal projection with one of the properties in Exercise 8.7 will be called 𝐽-positive and 𝐽-negative, respectively. Exercise 8.8 Try to prove the following. If 𝑃, 𝑄 are 𝐽-orthogonal projections such that 𝑃 is 𝐽-positive (𝐽-negative) and 𝑃 𝑄 = 𝑄, then 1. 𝑄𝑃 = 𝑄, 2. 𝑄 is 𝐽-positive (𝐽-negative), 3. ∥𝑄∥ ≤ ∥𝑃 ∥. Hint: Use Exercise 7.2, the representation (8.1) and Theorem 8.1.
Chapter 9
Spectral Properties and Reduction of 𝑱 -Hermitian Matrices
Here we start to study the properties of the eigenvalues and eigenvectors of 𝐽-Hermitian matrices. Particular attention will be given to the similarities and differences from (standard) Hermitian matrices. We begin with a list of properties which are more or less analogous to the ones of common Hermitian matrices. Theorem 9.1 Let 𝐴 be 𝐽-Hermitian. Then 1. 2. 3. 4.
If both 𝐴 and 𝐽 are diagonal, then 𝐴 is real. The spectrum of 𝐴 is symmetric with respect to the real axis. Any eigenvalue whose eigenvector is not 𝐽-neutral, is real. If 𝜆 and 𝜇 are eigenvalues and 𝜆 ∕= 𝜇, then the corresponding eigenvectors are 𝐽-orthogonal. 5. If a subspace 𝒳 is invariant under 𝐴, then so is its 𝐽-orthogonal companion. 6. The following are equivalent (i) There is a 𝐽-non-degenerate subspace, invariant under 𝐴. (ii) 𝐴 commutes with a non-trivial 𝐽-orthogonal projection. (iii) There is a 𝐽, 𝐽 ′ - unitary 𝑈 with 𝑈
−1
[
] 𝐴1 0 𝐴𝑈 = , 𝐽 ′ = diag(±1). 0 𝐴2
(9.1)
If the subspace from (i) is 𝐽-definite then 𝑈 can be chosen such that the matrix 𝐴1 in (9.1) is real diagonal and ] ±𝐼 0 𝐽 = , 𝐽2′ = diag(±1). 0 𝐽2′ ′
[
K. Veseli´ c, Damped Oscillations of Linear Systems, Lecture Notes in Mathematics 2023, DOI 10.1007/978-3-642-21335-9 9, © Springer-Verlag Berlin Heidelberg 2011
(9.2)
67
68
9 Spectral Properties and Reduction of 𝐽-Hermitian Matrices
Proof. The proofs of 1–5 are immediate and are omitted. To prove 6 let 𝒳 be 𝐽-non-degenerate. Then there is a 𝐽-orthogonal basis in 𝒳 , represented as the columns of 𝑈1 , so 𝑈1∗ 𝐽𝑈1 = 𝐽1 = diag(±1) and
𝐴𝑈1 = 𝑈1 𝑀
with
𝑀 = 𝐽1 𝑈1∗ 𝐽𝐴𝑈1
and with 𝑃 = 𝑈1 𝐽1 𝑈1∗ 𝐽 we have 𝑃 2 = 𝑃 , 𝑃 [∗] = 𝑃 and 𝐴𝑃 = 𝐴𝑈1 𝐽1 𝑈1∗ 𝐽 = 𝑈1 𝐽1 𝑈1∗ 𝐽𝐴𝑈1 𝐽1 𝑈 ∗ 𝐽 = 𝑃 𝐴𝑃. By taking [∗]-adjoints, we obtain 𝐴𝑃 = 𝑃 𝐴. Conversely, if 𝐴𝑈1 𝐽1 𝑈1∗ 𝐽 = 𝑈1 𝐽1 𝑈1∗ 𝐽𝐴, then by postmultiplying by 𝑈1 , 𝐴𝑈1 = 𝑈1 𝐽1 𝑈1∗ 𝐽𝐴𝑈1 , i.e. 𝒳 = ℛ(𝑈1 ) is invariant under 𝐴. Thus, (i) and (ii) are equivalent. Now, 𝑈 from (iii) is obtained by completing the columns of 𝑈1 to a 𝐽orthonormal basis, see Corollary 5.12. Finally, if the subspace is 𝐽-definite then by construction (9.2) will hold. Since 𝑈 −1 𝐴𝑈 is 𝐽 ′ -Hermitian the block 𝐴1 will be Hermitian. Hence there is a unitary 𝑉 such that 𝑉 −1 𝐴1 𝑉 is real diagonal. Now replace 𝑈 by [
] 𝑉 0 𝑈 =𝑈 . 0 𝐼 Q.E.D. Example 9.2 Let 𝐴 be 𝐽-Hermitian and 𝐴𝑢 = 𝜆𝑢 with [𝑢, 𝑢] = 𝑢∗ 𝐽𝑢 ∕= 0. Then the 𝐽- orthogonal projection along 𝑢, that is, onto the subspace spanned by 𝑢 is 𝑢𝑢∗ 𝐽 𝑃 = ∗ 𝑢 𝐽𝑢 and it commutes with 𝐴. Indeed, since 𝜆 is real we have 𝐴𝑃 = =
𝐴𝑢𝑢∗ 𝐽 𝑢𝑢∗ 𝐽 𝑢(𝐴𝑢)∗ 𝐽 = 𝜆 = 𝑢∗ 𝐽𝑢 𝑢∗ 𝐽𝑢 𝑢∗ 𝐽𝑢 𝑢𝑢∗ 𝐴∗ 𝐽 𝑢𝑢∗ 𝐽𝐴 = ∗ = 𝑃 𝐴. ∗ 𝑢 𝐽𝑢 𝑢 𝐽𝑢
9 Spectral Properties and Reduction of 𝐽-Hermitian Matrices
Also ∥𝑃 ∥2 = ∥𝑃 ∗ 𝑃 ∥ =
69
∥𝐽𝑢𝑢∗ 𝑢𝑢∗ 𝐽∥ ∥𝑢∗ 𝑢∥2 = , (𝑢∗ 𝐽𝑢)2 (𝑢∗ 𝐽𝑢)2
hence ∥𝑃 ∥ =
𝑢∗ 𝑢 . ∣𝑢∗ 𝐽𝑢∣
(9.3)
We call a 𝐽-Hermitian matrix 𝐴 𝐽, 𝐽 ′ -unitarily diagonalisable, if there is a 𝐽, 𝐽 ′ -unitary matrix 𝑈 such that the 𝐽 ′ -Hermitian matrix 𝐴′ = 𝑈 −1 𝐴𝑈 is diagonal. Note that 𝐽 ′ itself need not be diagonal but, if it is then the diagonal elements of 𝐴′ , i.e. the eigenvalues, must be real. The 𝐽, 𝐽 ′ -unitary block-diagonalisability is defined analogously. Example 9.3 We consider the one dimensional damped system 𝑚¨ 𝑥 + 𝑐𝑥˙ + 𝑘𝑥 = 0, 𝑚, 𝑐, 𝑘 > 0.
(9.4)
The phase-space matrix [ 𝐴=
] 0 𝜔 , −𝜔 −𝑑
is 𝐽-symmetric with
𝜔= [
𝐽=
√ 𝑘/𝑚, 𝑑 = 𝑐/𝑚
(9.5)
] 1 0 . 0 −1
Any 2 × 2 real 𝐽-orthogonal matrix (up to some signs) is of the form [ 𝑈 = 𝑈 (𝑥) =
] cosh 𝑥 sinh 𝑥 , sinh 𝑥 cosh 𝑥
(9.6)
with 𝑈 (𝑥)−1 = 𝑈 (−𝑥). We have 𝑈
−1
[
] 𝜔 sinh 2𝑥 + 𝑑 sinh2 𝑥 𝜔 cosh 2𝑥 + 𝑑2 sinh 2𝑥 𝐴𝑈 = . −𝜔 cosh 2𝑥 − 𝑑2 sinh 2𝑥 −𝜔 sinh 2𝑥 − 𝑑 cosh2 𝑥
Requiring the transformed matrix to be diagonal gives 2𝜔 tanh 2𝑥 = − =− 𝑑
√
4𝑘𝑚 , 𝑐2
(9.7)
70
9 Spectral Properties and Reduction of 𝐽-Hermitian Matrices
which is solvable if and only if 𝑐2 > 4𝑘𝑚. To better understand this, we will find the eigenvalues of 𝐴 directly: [ ] [ ] 𝑥 𝑥 𝐴 =𝜆 𝑦 𝑦 leads to 𝜔𝑦 = 𝜆𝑥 −𝜔𝑥 − 𝑑𝑦 = 𝜆𝑦 and finally to
𝑚𝜆2 + 𝑐𝜆 + 𝑘 = 0
which is the characteristic equation of the differential equation (9.4). The roots are √ −𝑐 ± 𝑐2 − 4𝑘𝑚 𝜆± = . (9.8) 2𝑚 According to the sign of the discriminant we distinguish three cases. 1. 𝑐2 > 4𝑘𝑚, the system is ‘overdamped’. There are two distinct real eigenvalues and the matrix 𝑈 in (9.6) exists with [ ] 𝜆+ 0 𝑈 −1 𝐴𝑈 = (9.9) 0 𝜆− 2. 𝑐2 < 4𝑘𝑚, the system is ‘weakly damped’. Here diagonalisation is impossible, but we may look for 𝑈 such that in 𝑈 −1 𝐴𝑈 the diagonal elements are equal, which leads to 𝑑 tanh 2𝑥 = − =− 2𝜔 and 𝑈
−1
√
[
𝑐2 0, i = 1, . . . , n i.e. Jp(A) is positive definite. In our example above we could choose p(λ) = (λ − 2)(λ + 1). For a general A take a J, J -unitary matrix U such that both J = U ∗ JU and A = U −1AU are diagonal (Corollary 10.2). Now, J p(A ) = U ∗ JU p(U −1 AU ) = U ∗ Jp(A)U is again positive definite by virtue of the Sylvester theorem. Q.E.D. To any J-Hermitian A with definite spectrum we associate the sequence s(A) = (n1 , . . . , nr , ±) characterising the r sign groups (in increasing order); the first r numbers carry their cardinalities whereas ± ∈ {−1, 1} indicates the first of the alternating signs. We shall call s(A) the sign partition of A or, equivalently, of the pair JA, J. The sign partition is obviously completely determined by any such p(λ) which we call the definitising polynomial. If a matrix with definite spectrum varies in a continuous way then the spectrum stays definite and the sign partition remains constant as long as the different sign groups do not collide. To prove this we need the following lemma. Lemma 10.4 Let I t → H(t) be a Hermitian valued continuous function on a closed interval I. Suppose that all H(t) are non-singular and that H(t0 ) is positive definite for some t0 ∈ I. Then all H(t) are positive definite. Proof. For any t1 there is an -neighbourhood in which the inertia of H(t) is constant. This is due to the continuity property of the eigenvalues. By the compactness the whole of I can be covered by a finite number of such neighbourhoods. This, together with the positive definiteness of H(t0 ) implies positive definiteness for all H(t). Q.E.D. Theorem 10.5 Let I t → A(t) be J-Hermitian valued continuous function on a closed real interval I such that 1. A(t0 ) has definite spectrum for some t0 ∈ I and 2. There are continuous real valued functions I t → fk (t), k = 1, . . . , p − 1 such that
76
10 Definite Spectra
(a) σ1 < f1 (t0 ) < σ2 < · · · < fp−1 (t0 ) < σp , where σ1 < σ2 < · · · < σp are the sign groups of A(t0 ) and (b) fk (t) ∩ σ(A(t)) = ∅ for all k and t. Then A(t) has definite spectrum for all t and s(A(t)) is constant in t. Proof. Let σ1 be, say, J-negative. Set pt (A(t)) = (A(t) − f1 (t)I)(A(t) − f2 (t)I)(A(t) − f3 (t)I) · · · then H(t) = Jpt (A(t)) satisfies the conditions of Lemma 10.4 and the statement follows. Q.E.D. From the proof of Theorem 10.3 it is seen that the definitising polynomial p can be chosen with the degree one less than the number of sign groups of A. In the case of just two such groups p will be linear and we can without loss of generality assume p(λ) as λ − μ. Then Jp(A) = JA − μJ and the matrix A (and also the matrix pair JA, J) is called J-definitisable or just definitisable. Any μ for which JA − μJ is positive definite is called a definitising shift. Of course, completely analogous properties are enjoyed by matrices for which JA − μJ is negative definite for some μ (just replace A by −A). Theorem 10.6 Let A be J-Hermitian. The following are equivalent: (i) A is definitisable. (ii) The spectrum of A is definite and the J-positive eigenvalues are larger then the J-negative ones. In this case the set of all definitising shifts form an open interval whose ends are eigenvalues; this is called the definiteness interval of A (or, equivalently, of the definitisable Hermitian pair JA, J. Proof. If JA − μJ is positive definite, then Ax = λx or, equivalently, (JA − μJ)x = (λ − μ)Jx implies x∗ (JA − μJ)x = (λ − μ)x∗ Jx. Thus, x is J-positive or J-negative according to whether λ > μ or λ < μ (μ itself is not an eigenvalue). Thus σ− (A) < μ < σ+ (A) where σ± (A) denote the set of J-positive/J-negative eigenvalues. Conversely, suppose σ− (A) < σ+ (A). The matrix A is J, J -unitarily diagonalisable: ⎡ ⎡ ⎤ ⎤ λ1 I1 1 I1 ⎢ ⎢ ⎥ ⎥ ∗ .. .. A = U −1 AU = ⎣ ⎦ , J = U JU = ⎣ ⎦, . . λp I
p Ip
10 Definite Spectra
77
where λ1 , . . . , λp are distinct eigenvalues of A and i ∈ {−1, 1}. Take any μ between σ− (A) and σ+ (A). Then ⎡ ⎢ J A − μJ = ⎣
⎤
1 (λ1 − μ)I1 ..
⎥ ⎦
.
(10.2)
p (λp − μ)Ip and this matrix is positive definite because the product i (λi − μ) is always positive. Now, J A − μJ = U ∗ JU U −1AU − μU ∗ JU = U ∗ (JA − μJ)U and JA−μJ is positive definite as well. The last assertion follows from (10.2). Q.E.D. The following theorem is a generalisation of the known fact for standard Hermitian matrices: the eigenvalues are ‘counted’ by means of the inertia. From now on we will denote the eigenvalues of a definitisable A as − + + λ− n− ≤ · · · ≤ λ1 < λ1 ≤ · · · ≤ λn+ .
(10.3)
Theorem 10.7 Let A be definitisable with the eigenvalues as in (10.3) and + + the definiteness interval (λ− 1 , λ1 ). For any λ > λ1 , λ ∈ σ(A) the quantity ι− (JA − λJ) equals the number of the J-positive eigenvalues less than λ (and similarly for J-negative eigenvalues). Proof. Since the definitisability is invariant under any J, J -unitary similarity and A is J, J -unitarily diagonalisable, we may assume A and J to be diagonal, i.e. − A = diag(λ+ J = diag(In+ , In− ). n+ , . . . , λn− ), Then + − − ι− (JA − λJ) = ι− (diag(λ+ n+ − λ, . . . , λ1 − λ, −(λ1 − λ), . . . , −(λn− − λ)))
and the assertion follows. Q.E.D. The following ‘local counting property’ may be useful in studying general J-Hermitian matrices. Theorem 10.8 Let A be J-Hermitian and let λ0 be a J-positive eigenvalue of multiplicity p. Take > 0 such that the open interval I = (λ0 − , λ0 + ) contains no other eigenvalues of A. Then for λ0 − < λ− < λ0 < λ+ < λ0 +
78
10 Definite Spectra
we have ι+ (JA − λ− J) = ι+ (JA − λ+ J) + p. Proof. By virtue of Proposition 10.1 and (10.1) and the fact that ι(JA − λJ ) = ι(U ∗ (JA − λJ)U ) = ι(JA − λJ) (Sylvester!) we may suppose that A, J have already the form A= Now
and
Ip 0 λ0 I 0 , J = , 0 A2 0 J2
I ∩ σ(A2 ) = ∅.
0 (λ0 − λ± )Ip JA − λ± J = 0 J2 A2 − λ± J2
ι+ (JA − λ− J) = p + ι+ (J2 A2 − λ− J2 ) = p + ι+ (J2 A2 − λ+ J2 ) = ι+ (JA − λ+ J).
Here we have used two trivial equalities ι+ (λ0 − λ− )Ip = p,
ι+ (λ0 − λ+ )Ip = 0
and the less trivial one ι+ (J2 A2 − λ+ J2 ) = ι+ (J2 A2 − λ− J2 ); the latter is due to the fact that for λ ∈ I the inertia of J2 A2 − λJ2 cannot change since this matrix is non-singular for all these λ. Indeed, by the known continuity of the eigenvalues of J2 A2 − λJ2 as functions of λ the matrix J2 A2 − λJ2 must become singular on the place where it would change its inertia, this is precluded by the assumption I ∩ σ(A2 ) = ∅. Q.E.D. Theorem 10.9 Let A be J-Hermitian and σ(A) negative. Then JA − μJ is positive definite, if and only if −JA−1 + J/μ is such. + Proof. Let JA − μJ be positive definite. Then λ− 1 < μ < λ1 . Obviously, the −1 matrix −A is J-Hermitian as well and it has the eigenvalues − + + 0 < −1/λ− n− ≤ · · · ≤ −1/λ1 < −1/μ < −1/λ1 ≤ · · · ≤ −1/λn+ , − + − where −1/λ+ 1 ≤ · · · ≤ −1/λn+ are J-positive and −1/λn− ≤ · · · ≤ −1/λ1 J-negative. The converse is proved the same way. Q.E.D.
Exercise 10.10 Try to weaken the condition σ(A) < 0 in the preceding theorem. Produce counterexamples.
10 Definite Spectra
79
We now consider the boundary of the set of definitisable matrices. Theorem 10.11 If JA − λJ is positive semidefinite and singular then we have the following alternative. 1. The matrix A is definitisable and λ lies on the boundary of the definiteness interval; in this case λ is a J-definite eigenvalue. 2. The matrix A is not definitisable or, equivalently, there is no λ = λ for which JA − λ J would be positive semidefinite. In this case all eigenvalues greater than λ are J-positive and those smaller than λ are J-negative whereas the eigenvalue λ itself is not J-definite. Proof. As in the proof of Theorem 10.6 we write Ax = λ x, λ > λ as (A − λI)x = (λ − λ)x, hence
x∗ (JA − λJ)x = (λ − λ)x∗ Jx,
which implies x∗ Jx ≥ 0. Now, x∗ Jx = 0 is impossible because x∗ (JA − λJ)x = 0 and the assumed semidefiniteness of JA − λJ would imply (JA − λJ)x = 0 i.e. Ax = λx. Thus, x∗ Jx > 0 and similarly for λ < λ. So, if λ itself is, say, J-positive, then the condition (ii) of Theorem 10.6 holds and the non-void definiteness interval lies left from λ. Finally, suppose that λ is not definite and that there is, say. λ < λ for which JA − λ J would be positive semidefinite. Then, as was shown above, λ would be J-positive which is impossible. Q.E.D. The definitisability of the pair JA, J can be expressed as [(A − λI)x, x] > 0 for some λ and all non-vanishing x. The eigenvalues of definitisable matrices enjoy extremal properties similar to those for standard Hermitian matrices. Consider the functional x → r(x) =
[Ax, x] x∗ JAx = , ∗ x Jx [x, x]
which is obviously real valued and defined on any non-neutral vector x. It will be called the Rayleigh quotient of A (or, which is the same, of the pair JA, J). Theorem 10.12 Let A ∈ Ξ n,n be J-Hermitian. Then A is definitisable, if and only if the values r+ = minn r(x), x∈Ξ x∗ Jx>0
r− = maxn r(x) x∈Ξ x∗ Jx 0 we have r(x) ≥ λ+ 1 while the equality is obtained on any eigenvector y for the eigenvalue λ+ 1 and nowhere else. Thus, the left equality in (10.4) is proved (and similarly for the other one). Conversely let r(x0 ) = r+ . For any vector h and any real ε we have r(x0 + εh) x∗ JAx0 + 2ε Re x∗0 JAh + ε2 h∗ JAh (x0 + εh)∗ JA(x0 + εh) = 0 ∗ Jh x∗ Jh ∗ (x0 + εh) J(x0 + εh) x∗0 Jx0 (1 + 2ε Re x∗0Jx0 + ε2 xh∗ Jx ) 0 0 0
2ε x∗ JAx0 x∗0 Jh = r+ + ∗ Re x∗0 JAh − 0 ∗ x0 Jx0 x0 Jx0 ε2 Re x∗0 JAh Re x∗0 Jh x∗0 Jh 2 ∗ + ∗ + 4(Re ) x JAx h∗ (JA − r+ J)h − 4 0 0 x0 Jx0 x∗0 Jx0 x∗0 Jx0 =
+O(ε3 ). This is a geometric series in ε which has a finite convergence radius (depending on h). Since this has a minimum at ε = 0, the coefficient at ε vanishes i.e.
10 Definite Spectra
81
Re(x∗0 (JA − r+ J)h) = 0 and since h is arbitrary we have JAx0 = r+ Jx0 . Using this we obtain r(x0 + εh) =
ε2 h∗ (JA − r+ J)h + O(ε3 ). x∗0 Jx0
From this and the fact that x0 is the minimum point it follows h∗ (JA − r+ J)h ≥ 0 and this is the positive semidefiniteness of JA − r+ J. The same for JA − r− J is obtained analogously. By Theorem 10.11 this implies that A is definitisable and that (r− , r+ ) is its definiteness interval. Q.E.D. In contrast to the standard Hermitian case the ‘outer’ boundary of the spectrum of a definitisable A is no extremum for the Rayleigh quotient. As an example take 2 0 1 0 A= , J= 0 −1 0 −1
cosh φ , x= sinh φ
and
then x∗ Jx = 1 and
r(x) = 1 + cosh2 φ,
which is not bounded from above. However, all eigenvalues can be obtained by minimax formulae, similar to those for the standard Hermitian case. We will now derive these formulae. We begin with a technical result, which is nonetheless of independent interest. Lemma 10.13 (The interlacing property) Let the matrix A be J-Hermitian and definitisable and partitioned as A=
A11 A12 A21 A22
and, accordingly, J =
J1 0 . 0 J2
(10.5)
Here A11 is square of order m. Then A11 is J1 -Hermitian and definitisable and its definiteness interval contains that of A. Denote by (10.3) the eigenvalues of A and by − + + μ− m− ≤ · · · ≤ μ1 < μ1 ≤ · · · ≤ μm+ ,
those of A11 . Then
m+ + m− = m,
+ + λ+ k ≤ μk ≤ λk+n+ −m+ − − λ− k ≥ μk ≥ λk+n− −m−
(here an inequality is understood as void whenever an index exceeds its range).
82
10 Definite Spectra
Proof. If JA − μJ is Hermitian positive definite then so is the submatrix J1 A11 − μJ1 . By the J-Hermitian property of A we have A21 = J2 A∗12 J1 hence J1 A12 J1 A11 − μJ1 JA − μJ = A∗12 J1 J2 A22 − μJ2 J A − μJ1 0 Z(μ)∗ = Z(μ) 1 11 0 W (μ)
with Z(μ) = and
0 Im , A∗12 J1 (J1 A11 − μJ1 )−1 In−m
W (μ) = J2 A22 − μJ2 − A∗12 J1 (J1 A11 − μJ1 )−1 J1 A12 .
By the Sylvester inertia theorem, ι± (JA − μJ) = ι± (J1 A11 − μJ1 ) + ι± (W (μ)), hence ι− (J1 A11 − μJ1 ) ≤ ι− (JA − μJ) ≤ ι− (J1 A11 − μJ1 ) − n − m.
(10.6)
+ Assume now μ+ k < λk for some k. Then there is a μ such that J1 A11 − μJ1 + is non-singular and μ+ k < μ < λk . By Theorem 10.7 we would have
ι− (J1 A11 − μJ1 ) ≥ k,
ι− (JA − μJ) < k
which contradicts the first inequality in (10.6). Similarly, μ+ k > μ > λ+ would imply k+n+ −m+ ι− (J1 A11 − μJ1 ) ≥ k,
ι− (JA − μJ) < k
which contradicts the second inequality in (10.6). Q.E.D. Theorem 10.14 For a definitisable J-Hermitian A the following minimax formulae hold: max λ± k = min ± ± Sk
x∈S k x∗ Jx=0
x∗ JAx [Ax, x] max = min , ± ± x∗ Jx [x, x] x∈S Sk k
(10.7)
[x,x]=0
where Sk± ⊆ Ξ n is any k-dimensional J-positive/J-negative subspace. Proof. Let Sk+ be given and let u1 , . . . , un be a J-orthonormal basis of Ξ n such that u1 , . . . , uk span Sk+ . Set
10 Definite Spectra
83
U + = u1 · · · u k ,
U = u1 · · · un = U + U .
Then U is J,J -unitary with
J1 0 J = 0 J2
Ik 0 = , 0 J2
and
A =U
−1
J2 = diag(±1)
A11 A12 AU = A21 A22
is J -Hermitian: A11 = A∗ 11 ,
A21 = J2 A∗ 12 ,
A22 = J2 A∗ 22 J2 .
Moreover, A and A have the same eigenvalues and the same definiteness interval. By Lemma 10.13 we have the interlacing of the corresponding eigenvalues + + λ+ k = 1, . . . m+ k ≤ μk ≤ λk+n+ −m+ , + (the eigenvalues μ− k are lacking). The subspace Sk consists of the vectors
y U , 0
y ∈ Ξk
thus for arbitrary x ∈ Sk+ we have x∗ JAx y ∗ A11 y = . ∗ x Jx y∗y Since A11 is (standard) Hermitian, μ+ k = max y=0
y ∗ A11 y x∗ JAx = max ≥ λ+ k ∗ Jx + y∗ y x x∈S k x=0
and since Sk+ is arbitrary it follows inf max Sk+
+ x∈S k x=0
x∗ JAx ≥ λ+ k. x∗ Jx
Now choose Sk+ as the linear span of J-orthonormal eigenvectors v1 , . . . , vk , + + belonging to the eigenvalues λ+ 1 , . . . , λk of A. Then any x ∈ Sk is given as x = α1 v1 + · · · + αk vk
84
10 Definite Spectra
T and, with α = α1 · · · αk , k
max +
x∈S k x=0
x∗ JAx + = max |αj |2 λ+ j = λk . α∗ α=1 x∗ Jx j=1
This proves the assertion for J-positive eigenvalues (the other case is analogous). Q.E.D. Remark 10.15 In the special case J = I Theorem 10.14 gives the known minimax formulae for a Hermitian matrix A of order n: λk = min max Sk
x∈Sk x=0
x∗ Ax , x∗ x
where Sk is any k-dimensional subspace and λk are non-decreasingly ordered eigenvalues of A. Now we can prove the general formula (2.8): by setting −∗ M = L2 L∗2 and A = L−1 2 KL2 we obtain max x∈Sk x=0
x∗ Ax = x∗ x
max
−∗ y∈L2 Sk y=0
y ∗ Ky y∗ M y
and L−∗ 2 Sk again varies over the set of all k-dimensional subspaces. This proves the ‘min max’ part of (2.8), the other part is obtained by considering the pair −K, M . Note, however, that for a general J-Hermitian matrix no ‘max min’ variant in the formulae (10.7) exists. The following extremal property could be useful for computational purposes. Theorem 10.16 Let A be J-Hermitian. Consider the function X → Tr(X ∗ JAX)
(10.8)
defined on the set of all J, J1 -isometries X ∈ Ξ n,m for a fixed symmetry J1 = diag(Im+ , −Im− ),
m± ≤ n± .
If A is definitisable with the eigenvalues (10.3) then the function (10.8) takes its minimum on any X of the form
X = X+ X− , where R(X ± ) is the subspace spanned by m± J-orthonormal eigenvectors for ± the eigenvalues λ± 1 , . . . , λm± of A, respectively – and nowhere else.
10 Definite Spectra
85
Proof. Without loss of generality we may suppose J to be of the form J=
In+ 0 . 0 −In−
Indeed, the transition from a general J to the one above is made by unitary similarity as described in Remark 5.1. The trace function (10.8) is invariant under these transformations. Also, by virtue of the Sylvester inertia theorem, applied to X ∗ JX = J1 , we have m+ ≤ n + ,
m− ≤ n− .
First we prove the theorem for the case m± = n± i.e. the matrices X are square and therefore J-unitary. Since A is definitisable we may perform a J-unitary diagonalisation U −1 AU = Λ,
Λ = diag(Λ+ , Λ− ),
± Λ± = diag(λ± 1 , . . . λn± ),
U ∗ JU = J.
Now Tr(X ∗ JAX) = Tr(X ∗ JU ΛU −1X) = Tr(X ∗ U −∗ JΛU −1 X) = Tr(Y ∗ JΛY ), where Y = U −1 X again varies over the set of all J-unitaries. By using the representation (6.2) we obtain = TrY ∗ JΛY TrX ∗ JAX √ = TrH (W ) JΛH (W ) √ = Tr( I + W W ∗ Λ+ I + W W ∗ − W Λ− W ∗ ) √ √ +Tr(W ∗ Λ+ W − I + W ∗ W Λ− I + W ∗ W ) = t0 + 2(TrW W ∗ Λ+ − TrW ∗ W Λ− ) = t0 + 2[TrW W ∗ (Λ+ − μI) + TrW ∗ W (μI − Λ− )], where μ is a definitising shift and t0 = TrΛ+ − TrΛ− . Since Λ+ − μI and μI − Λ− are positive definite it follows Tr(X ∗ JAX) = TrY ∗ JΛY = t0 + 2Tr (W ∗ (Λ+ − μI) W ) + 2Tr (W (μI − Λ− ) W ∗ ) ≥ t0 ,
(10.9)
moreover, if this inequality turns to equality then W = 0, hence H(W ) = I and the corresponding X reads X = U diag(V1 , V2 ),
86
10 Definite Spectra
V1 , V2 from (6.2). This is exactly the form of the minimising X in the statement of the theorem. ¯ as We now turn to TrX ∗ JAX with a non-square X. Take C = (X X) a completion of any J, J1 -isometry X such that the columns of C form a J-orthonormal basis, that is, C ∗ JC = diag(J1 , J2 ) = J is diagonal (note that the columns of X are J-orthonormal). Then the matrix A1 = C −1 AC is J -Hermitian and we have A1 =
¯ J1 X ∗ JAX J1 X ∗ JAX ¯ , ¯ ∗ JAX J2 X ¯ ∗ JAX J2 X
J A1 = C ∗ JAC =
¯ X ∗ JAX X ∗ JAX ∗ ∗ ¯ ¯ ¯ X JAX X JAX
and the eigenvalues of A1 are given by (10.3). Then by Lemma 10.13 the ∗ eigenvalues μ± k of the J1 -Hermitian matrix J1 X JAX satisfy + λ+ k ≤ μk ,
− λ− k ≥ μk .
Using this and applying the first part of the proof for the matrix J1 X ∗ JAX we obtain TrX ∗ JAX ≥
m+ i=1
μ+ i −
m− j=1
μ− j ≥
m+ i=1
λ+ i −
m− j=1
λ− j .
This lower bound is attained, if X is chosen as in the statement of the theorem, that is, if AX ± = X ± Λ± ,
± Λ± = diag(λ± 1 , . . . , λm± ).
It remains to determine the set of all minimisers. If X is any minimiser then we may apply the formalism of Lagrange with the Lagrange function L = Tr(X ∗ JAX) − Tr(Γ (X ∗ JX − J1 )), where the Lagrange multipliers are contained in the Hermitian matrix Γ of order m. Their number equals the number of independent equations in the isometry constraint X ∗ JX = J1 . By setting the differential of L to zero we obtain AX = XΓ and by premultiplying by X ∗ J,
10 Definite Spectra
87
X ∗ JAX = J1 Γ. Hence Γ commutes with J1 : Γ = diag(Γ + , Γ − ). This is the form of the minimiser, stated in the theorem. Q.E.D. When performing diagonalisation the question of the condition number of the similarity matrix naturally arises. We have Proposition 10.17 Let A be J-Hermitian and have a definite spectrum. Then all J, J - unitaries that diagonalise it with any diagonal J have the same condition number. Proof. Suppose U1−1 AU1 = Λ1 , U1∗ JU1 = J1 U2−1 AU2 = Λ2 , U2∗ JU2 = J2 , where Λ1 , Λ2 are diagonal and J1 , J2 are diagonal matrices of signs. The J-definiteness means that there exist permutation matrices Π1 , Π2 such that ⎡ ⎢ ⎢ Π1T J1 Π1 = Π2T J2 Π2 = ⎢ ⎣
1 In1
⎡ ⎢ ⎢ Π1T Λ1 Π1 = Π2T Λ2 Π2 = ⎢ ⎣
⎤ ⎥ ⎥ ⎥ =: J 0 , ⎦
2 In2 ..
. p Inp
λ1 In1
⎥ ⎥ ⎥ =: Λ, ⎦
λ2 In2 ..
⎤
. λp Inp
here i ∈ {−1, 1} are the signs of the corresponding subspaces and λ1 , . . . , λp are the distinct eigenvalues of A with the multiplicities n1 , . . . , np . Then both U˜1 = U1 Π1 and U˜2 = U2 Π2 are J, J 0 -unitary and ˜1 = Λ = U ˜ −1 AU ˜2 ˜ −1 AU U 1 2
(10.10)
or, equivalently, ΛV = V Λ, ˜ −1 U ˜2 is J 0 -unitary. Since the eigenvalues λ1 , . . . , λp where the matrix V = U 1 are distinct, (10.10) implies J 0V = V J 0
88
10 Definite Spectra
that is, V is unitary and ˜2 Π2T = U ˜1 V Π2T = U1 Π1 V Π2T , U2 = U where the matrix Π1 V Π2T is unitary, so U2 and U1 have the same condition numbers. Q.E.D. Note that the previous theorem applies not only to the standard, spectral norm but to any unitarily invariant matrix norm like e.g. the Euclidian norm. Exercise 10.18 Show that the ‘if ’ part of the Theorem 10.12 remains valid, if the symbols min/max in (10.4) are substituted by inf/sup. Exercise 10.19 Try to estimate the condition of the matrix X in (10.9), if the difference Tr(JA) − t0 is known.
Chapter 11
General Hermitian Matrix Pairs
Here we briefly overview the general eigenvalue problem Sx = λT x with two Hermitian matrices S, T and show how to reduce it to the case of a single J-Hermitian matrix A. The important property, characterised by Theorem 10.14, was expressed more naturally in terms of the Hermitian matrix pair JA, J, than in terms of the single J-Hermitian matrix A. In fact, the eigenvalue problem Ax = λx can be equivalently written as JAx = λJx. More generally, we can consider the eigenvalue problem Sx = λT x, (11.1) where S and T are Hermitian matrices and the determinant of S−λT does not identically vanish. Such pairs are, in fact, essentially covered by our theory. Here we outline the main ideas. If T is non-singular, then we may apply the decomposition (5.8) to the matrix T , replace there G by G|α|1/2 and set J = sign(α), then T = GJG∗ ,
(11.2)
where G is non-singular and J = diag(±1). By setting y = G∗ x, (11.1) is equivalent to Ay = λy, A = JG−1 SG−∗ , (11.3) where A is J-Hermitian. Also, det(A − λI) =
det(S − λT ) . det T
To any invariant subspace relation AY = Y Λ
K. Veseli´ c, Damped Oscillations of Linear Systems, Lecture Notes in Mathematics 2023, DOI 10.1007/978-3-642-21335-9 11, © Springer-Verlag Berlin Heidelberg 2011
89
90
11 General Hermitian Matrix Pairs
with Y injective, there corresponds SX = T XΛ, Y = G∗ X and vice versa. The accompanying indefinite scalar product is expressed as y ∗ Jy = x∗ T x with y = G∗ x, y = G∗ x . Thus, all properties of J-Hermitian matrices studied in the two previous chapters can be appropriately translated into the language of matrix pairs S, T with non-singular T and vice versa. The non-singularity of T will be tacitly assumed in the following. (If T is singular but there is a real μ such that T − μS is non-singular then everything said above can be done for the pair S, T − μS. Now, the equations Sx = λ(T − μS)x,
Sx =
λ Tx 1 − λμ
are equivalent. Thus the pair S, T − μS has the same eigenvectors as S, T while the eigenvalues λ of the former pair are transformed into λ/(1 − λμ) for the latter one.) The terms ‘definite eigenvalues’ and ‘definitisability’ carry over in a natural way to the general pair S, T by substituting the matrix JA for S and the symmetry J for T . Theorems 10.6–10.16 can be readily formulated and proved in terms of the matrix pair S, T . This we leave to the reader and, as an example, prove the following Theorem 11.1 Let S, T be definitisable and ι(T ) = (n+ , n− , 0). Then there is a Ψ such that Ψ ∗ T Ψ = J = diag(In+ , −In− ), + − − Ψ ∗ SΨ = diag(λ+ n+ , . . . , λ1 , −λ1 , . . . , −λn− ) + − − with λ+ n+ ≥ · · · ≥ λ1 ≥ λ1 ≥ · · · ≥ λn− . Moreover, for m± ≤ n± and J1 = diag(Im+ , −Im− ) the function
Φ → Tr(Φ∗ SΦ), defined on the set of all Φ with Φ∗ T Φ = J1 takes its minimum
n+
+ k=1 λk
−
n−
k=1
λ− k on any
Φ = Φ+ Φ− ,
(11.4)
11 General Hermitian Matrix Pairs
91
where R(Φ± ) is spanned by J-orthonormal eigenvectors belonging to the ± eigenvalues λ± 1 , . . . , λn± – and nowhere else. Proof. In (11.2) we may take J with diagonals ordered as above. Then with A = JG−1 SG−∗ for any μ we have JA − μJ = G∗ (S − μT )G so by the non-singularity of G the definitisability of the pair S, T is equivalent to that of JA, J that is, of the J-Hermitian matrix A. By Corollary 10.2 there is a J-unitary U such that + − − U −1 AU = diag(λ+ n+ , . . . , λ1 , λ1 , . . . , λn− ).
By setting Ψ = G−∗ U we have the diagonalisation + − Ψ ∗ SΨ = U ∗ JAU = JU −1 AU = diag(λ+ n+ , . . . , λ1 , −λ1 , . . . , −λn− ).
Now set X = G∗ Φ, then the ‘T, J-isometry condition’ (11.4) is equivalent to X ∗ JX = J1 and we are in the conditions of Theorem 10.16. Thus, all assertions of our theorem follow immediately from those of Theorem 10.16. Q.E.D. Remark 11.2 A definitisable pair S, T with T non-singular, can be diagonalised by applying the decomposition (2.1) to K = T and M = S − μT , μ a definitising shift: Φ∗ (S − μT )Φ = I,
Φ∗ T Φ = diag(α1 , . . . , αn ).
Then Ψ = Φ diag(|α1 |−1/2 , . . . , |αn |−1/2 ) satisfies Ψ ∗ T Ψ = J = diag(sign(αj )) and Ψ ∗ SΨ = Ψ ∗ (S − μT + μT )Ψ = J diag(
1 1 + μ, . . . , + μ). α1 αn
The desired order of the signs in J can be obtained by appropriately permuting the columns of Ψ . ˙ The system (1.1) can also be represented as follows. We set x1 = x, x2 = x, then (1.1) goes over into d x 0 x1 =S 1 + T dt x2 x2 −f (t)
(11.5)
92
with
11 General Hermitian Matrix Pairs
K 0 0 K T = , S= , 0 −M K C
(11.6)
where the matrix T is non-singular, if both K and M are such. The representations (3.2), (3.3) and (11.5), (11.6) are connected by the formulae (11.2) and (11.3), where L1 0 . G= 0 L2
The choice of the representation is a matter of convenience. A reason for choosing the ‘normalised representation’ (3.2), (3.3) is the physically natural phase space with its energy norm. Another reason is that here we have to do with a single matrix A while J is usually stored as a sequence of signs which may have computational advantages when dealing with dense matrices. Also the facts about condition numbers are most comfortably expressed in the normalised representation. On the other hand, if we have to deal with large sparse matrices S, T , then the sparsity may be lost after the transition to JA, J and then the ‘non-normalised’ representation (11.5), (11.6) is more natural in numerical computations.
Chapter 12
Spectral Decomposition of a General 𝑱 -Hermitian Matrix
A general 𝐽-Hermitian matrix 𝐴 may not have a basis of eigenvectors. In this chapter we describe the reduction to a block-diagonal form by a similarity transformation 𝑆 −1 𝐴𝑆. We pay particular attention to the problem of the condition number of the transformation matrix 𝑆 which is a key quantity in any numerical manipulation. The formal way to solve the initial value problem 𝑦˙ = 𝐴𝑦,
𝑦(0) = 𝑦0
is to diagonalise 𝐴:1 𝑆 −1 𝐴𝑆 = diag(𝜆1 , . . . , 𝜆𝑛 ). Then Set
(12.1)
𝑆 −1 𝑒𝐴𝑡 𝑆 = diag(𝑒𝜆1 𝑡 , . . . , 𝑒𝜆𝑛 𝑡 ). [ ] 𝑆 = 𝑠1 ⋅ ⋅ ⋅ 𝑠𝑛 .
Then the solution is obtained in two steps: • Compute 𝑦0′ from the linear system
• Set
𝑆𝑦0′ = 𝑦0
(12.2)
′ ′ 𝑦 = 𝑒𝜆1 𝑡 𝑦0,1 𝑠1 + ⋅ ⋅ ⋅ + 𝑒𝜆𝑛 𝑡 𝑦0,𝑛 𝑠𝑛 .
(12.3)
This method is not viable, if 𝐴 cannot be diagonalised or if the condition number of the matrix 𝑆 is too high. The latter case will typically occur when the matrix 𝐴 is close to a non-diagonalisable one. High condition of 𝑆 will
1 In
this chapter we will consider all matrices as complex.
K. Veseli´ c, Damped Oscillations of Linear Systems, Lecture Notes in Mathematics 2023, DOI 10.1007/978-3-642-21335-9 12, © Springer-Verlag Berlin Heidelberg 2011
93
94
12 Spectral Decomposition of a General 𝐽-Hermitian Matrix
spoil the accuracy of the solution of the linear system (12.2) which is a vital step in computing the solution (12.3). Things do not look better, if we assume 𝐴 to be 𝐽-Hermitian – with the exception of definite spectra as shown in Chap. 10. An example of a nondiagonalisable 𝐽-Hermitian 𝐴 was produced in (9.5) with 𝑑 = 2𝜔 (critical damping). Instead of the ‘full diagonalisation’ (12.1) we may seek any reduction to a block-diagonal form 𝑆 −1 𝐴𝑆 = diag(𝐴′1 , . . . , 𝐴′𝑝 ). so the exponential solution is split into ′
′
𝑆 −1 𝑒𝐴𝑡 𝑆 = diag(𝑒𝐴1 𝑡 , . . . , 𝑒𝐴𝑝 𝑡 ). A practical value of block-diagonalisation lies in the mere fact that reducing the dimension makes computation easier. So the ideal would be to obtain as small sized diagonal blocks as possible while keeping the condition number of 𝑆 reasonably low. What we will do here is to present the spectral reduction, that is, the reduction in which the diagonal blocks have disjoint spectra. We will pay particular attention to matrices 𝐴 that are 𝐽-Hermitian. The approach we will take will be the one of complex contour of analytic functions of the complex variable 𝜆. We will freely use the general properties of matrix-valued analytic functions which are completely analogous to those in the scalar case. One of several equivalent definitions of analyticity is the analyticity of the matrix elements.2 Our present considerations will be based on one such function, the resolvent : 𝑅(𝜆) = (𝜆𝐼 − 𝐴)−1 , which is an analytic, more precisely, rational function in 𝜆 as revealed by the Cramer-rule formula 𝐴adj (𝜆) 𝑅(𝜆) = , det(𝜆𝐼 − 𝐴) where the elements of 𝐴adj (𝜆) are some polynomials in 𝜆. The fundamental formula of the ‘analytic functional calculus’ is ∫ 1 𝑓 (𝐴) = 𝑓 (𝜆)(𝜆𝐼 − 𝐴)−1 𝑑𝜆, 2𝜋𝑖 𝛤 2 The
(12.4)
non-commutativity of matrix multiplication carries only minor, mostly obvious, changes in the standard formulae of the calculus: (𝐴(𝜆)𝐵(𝜆))′ = 𝐴′ (𝜆)𝐵(𝜆) + 𝐴(𝜆)𝐵 ′ (𝜆), (𝐴(𝜆)−1 )′ = −𝐴(𝜆)−1 𝐴′ (𝜆)𝐴(𝜆)−1 and the like.
12 Spectral Decomposition of a General 𝐽-Hermitian Matrix
95
where 𝑓 (𝜆) is any analytic function in a neighbourhood 𝒪 of 𝜎(𝐴) (𝒪 need not be connected) and 𝛤 ⊆ 𝒪 is any contour surrounding 𝜎(𝐴). The map 𝑓 → 𝑓 (𝐴) is continuous and has the properties (𝛼𝑓1 + 𝛽𝑓2 )(𝐴) = 𝛼𝑓1 (𝐴) + 𝛽𝑓2 (𝐴)
(12.5)
(𝑓1 ⋅ 𝑓2 )(𝐴) = 𝑓1 (𝐴)𝑓2 (𝐴) (⋅ is the pointwise multiplication of functions) (12.6) (𝑓1 ∘ 𝑓2 )(𝐴) = 𝑓1 (𝑓2 (𝐴)) (∘ is the composition of functions) 1(𝐴) = 𝐼, where 1(𝜆) = 1
(12.7)
𝝀(𝐴) = 𝐴, where 𝝀(𝜆) = 𝜆
(12.8)
𝜎(𝑓 (𝐴)) = 𝑓 (𝜎(𝐴)). All these properties are readily derived by using the resolvent equation 𝑅(𝜆) − 𝑅(𝜇) = (𝜆 − 𝜇)𝑅(𝜆)𝑅(𝜇)
(12.9)
as well as the resulting power series 𝑅(𝜆) =
∞ ∑ 𝐴𝑘 , 𝜆𝑘+1
(12.10)
𝑘=0
valid for ∣𝜆∣ > ∥𝐴∥. We just sketch some of the proofs: (12.5) follows immediately from (12.4), for (12.6) use (12.9) where in computing 𝑓1 (𝐴)𝑓2 (𝐴) =
1 2𝜋𝑖
∫ 𝛤1
𝑓1 (𝜆)(𝜆𝐼 − 𝐴)−1 𝑑𝜆
∫ 𝛤2
𝑓1 (𝜇)(𝜇𝐼 − 𝐴)−1 𝑑𝜇,
the contour 𝛤2 is chosen so that it is contained in the 𝛤1 -bounded neighbourhood of 𝜎(𝐴). Equation (12.7) follows from (12.10) whereas (12.8) follows from (12.7) and the identity 𝜆(𝜆𝐼 − 𝐴)−1 = (𝜆𝐼 − 𝐴)−1 − 𝐼. The properties (12.5)–(12.8) imply that the definition (12.4) coincides with other common definitions of matrix functions like matrix polynomials, rational functions and convergent power series, provided that the convergence disk contains the whole spectrum. For example, 𝛼0 𝐼 + ⋅ ⋅ ⋅ + 𝛼𝑝 𝐴𝑝 =
1 2𝜋𝑖
∫ 𝛤
(𝛼0 + ⋅ ⋅ ⋅ + 𝛼𝑝 𝜆𝑝 )(𝜆𝐼 − 𝐴)−1 𝑑𝜆,
−1
(𝜇𝐼 − 𝐴)
1 = 2𝜋𝑖
∫ 𝛤
1 (𝜆𝐼 − 𝐴)−1 𝑑𝜆, 𝜇−𝜆
(12.11) (12.12)
96
12 Spectral Decomposition of a General 𝐽-Hermitian Matrix
𝑒𝐴 = 1 𝐴 = 2𝜋𝑖 𝛼
∫ 𝛤
∞ ∑ 𝐴𝑘 𝑘=0
𝑘!
𝛼
=
𝜆 (𝜆𝐼 − 𝐴)
1 2𝜋𝑖
−1
∫ 𝛤
𝑒𝜆 (𝜆𝐼 − 𝐴)−1 𝑑𝜆,
(12.13)
∑ (𝛼) 𝑑𝜆 = (𝐴 − 𝐼)𝑘 . (12.14) 𝑘 𝑘
Here 𝐴𝛼 depends on the choice of the contour 𝛤 and the last equality holds for ∥𝐼 − 𝐴∥ < 1 (it describes the branch obtained as the analytic continuation starting from 𝐴 = 𝐼, 𝐴𝛼 = 𝐼). We will always have to work with fractional powers of matrices without non-positive real eigenvalues. So, by default the contour 𝛤 is chosen so that it does not intersect the non-positive real axis. Exercise 12.1 Find conditions for the validity of the formula 𝐴𝛼 𝐴𝛽 = 𝐴𝛼+𝛽 . Another common expression for 𝑓 (𝐴) is obtained by starting from the obvious property ∫ 1 𝑓 (𝑆 −1 𝐴𝑆) = 𝑓 (𝜆)𝑆 −1 (𝜆𝐼 − 𝐴)−1 𝑆𝑑𝜆 = 𝑆 −1 𝑓 (𝐴)𝑆. (12.15) 2𝜋𝑖 𝛤 Now, if 𝑆 diagonalises 𝐴 i.e. 𝑆 −1 𝐴𝑆 = diag(𝜆1 , . . . , 𝜆𝑛 ) then (12.15) yields 𝑓 (𝐴) = 𝑆 diag(𝑓 (𝜆1 ), . . . , 𝑓 (𝜆𝑛 ))𝑆 −1 . Similarly, if 𝑆 block-diagonalises 𝐴 i.e. 𝑆 −1 𝐴𝑆 = diag(𝐴′1 , . . . , 𝐴′𝑝 ) then (12.15) yields 𝑓 (𝐴) = 𝑆 diag(𝑓 (𝐴′1 ), . . . , 𝑓 (𝐴′𝑝 ))𝑆 −1 . Exercise 12.2 Show that the map 𝐴 → 𝑓 (𝐴) is continuous. Hint: for 𝐴ˆ = 𝐴 + 𝛿𝐴 show that the norm of the series ˆ − 𝑓 (𝐴) = 𝑓 (𝐴)
∫ ∞ ∑ ( )𝑘 1 𝑓 (𝜆)(𝜆𝐼 − 𝐴)−1 𝛿𝐴(𝜆𝐼 − 𝐴)−1 𝑑𝜆 2𝜋𝑖 𝛤
(12.16)
𝑘=1
can be made arbitrarily small, if ∥𝛿𝐴∥ is small enough. Other fundamental functions 𝑓 are those which have their values in {0, 1}, the corresponding 𝑃 = 𝑓 (𝐴) are obviously projections. Obviously, any such projection (called spectral projection) is given by 1 𝑃 = 𝑃𝜎 = 2𝜋𝑖
∫ 𝛤
(𝜆𝐼 − 𝐴)−1 𝑑𝜆,
(12.17)
12 Spectral Decomposition of a General 𝐽-Hermitian Matrix
97
where 𝛤 separates the subset 𝜎 ⊆ 𝜎(𝐴) from the rest of 𝜎(𝐴). Also obvious are the relations 𝑃𝜎 𝑃𝜎′ = 𝑃𝜎′ 𝑃𝜎 = 0, 𝑃𝜎 + 𝑃𝜎′ = 𝑃𝜎′ ∪𝜎 whenever 𝜎 ∩ 𝜎 ′ = ∅. Thus, any partition 𝜎1 , . . . , 𝜎𝑝 of 𝜎(𝐴) gives rise to a decomposition of the identity 𝑃𝜎1 , . . . , 𝑃𝜎𝑝 , commuting with 𝐴. Taking a matrix 𝑆 whose columns contain the bases of ℛ(𝑃𝜎1 ), . . . , ℛ(𝑃𝜎𝑝 ) we obtain block-diagonal matrices ⎡ ′ ⎤ 𝐴1 ⎢ ⎥ ′ .. 𝐴′ = 𝑆 −1 𝐴𝑆 = ⎣ (12.18) ⎦ , 𝜎(𝐴𝑗 ) = 𝜎𝑗 , . 𝐴′𝑝 ⎡ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ −1 𝑆 𝑃𝜎𝑗 𝑆 = ⎢ ⎢ ⎢ ⎢ ⎢ ⎣
0
⎤
⎥ ⎥ ⎥ ⎥ 0 ⎥ ⎥ (12.19) 𝐼𝑛𝑗 ⎥ = 𝑃𝑗0 . ⎥ ⎥ 0 ⎥ .. ⎥ . ⎦ 0 Here 𝑛𝑗 is the dimension of the space ℛ(𝑃𝜎𝑗 ). If 𝜎𝑘 is a single point: 𝜎𝑘 = {𝜆𝑘 } then the space ℛ(𝑃𝜎𝑘 ) is called the root space belonging to 𝜆 ∈ 𝜎(𝐴) and is denoted by ℰ𝜆 . If all ℛ(𝑃𝜎𝑗 ) are root spaces then we call (12.18) the spectral decomposition of 𝐴. Here we see the advantage of the contour integrals in decomposing an arbitrary matrix. The decomposition (12.18) and in particular the spectral decomposition are stable under small perturbations of the matrix 𝐴. Indeed, the formula (12.16) can be applied to the projections in (12.17): they change continuously, if 𝐴 changes continuously. The same is then the case with the subspaces onto which they project. ..
.
Exercise 12.3 Show that 1 Tr𝑃 = Tr 2𝜋𝑖
∫ 𝛤
(𝜆𝐼 − 𝐴)−1 𝑑𝜆
(12.20)
equals the number of the eigenvalues of 𝐴 within 𝛤 together with their multiplicities whereas ∫ ˆ = Tr 1 𝜆 𝜆(𝜆𝐼 − 𝐴)−1 𝑑𝜆 2𝜋𝑖 𝛤 equals their sum. Hint: reduce 𝐴 to the triangular form.
98
12 Spectral Decomposition of a General 𝐽-Hermitian Matrix
Exercise 12.4 Show that ℛ(𝑃𝜎 ) and ℛ(𝑃𝜎 ) have the same dimensions. Hint: use (12.20). Exercise 12.5 Show that for a real analytic function 𝑓 and a real matrix 𝐴 the matrix 𝑓 (𝐴) is also real. Exercise 12.6 Let 𝐴, 𝐵 be any matrices such that the products 𝐴𝐵 and 𝐵𝐴 exist. Then 𝜎(𝐴𝐵) ∖ {0} = 𝜎(𝐵𝐴) ∖ {0} (12.21) and
𝐴𝑓 (𝐵𝐴) = 𝑓 (𝐴𝐵)𝐴.
Hint: Use (12.21) and the identity 𝐴(𝜆𝐼 − 𝐵𝐴)−1 = (𝜆𝐼 − 𝐴𝐵)−1 𝐴 (here the two identity matrices need not be of the same order). Exercise 12.7 Show that any simple real eigenvalue of a 𝐽-Hermitian 𝐴 is 𝐽-definite. Proposition 12.8 Let ℰ𝜆1 be the root space belonging to any eigenvalue 𝜆1 of 𝐴. Then ℰ𝜆1 = {𝑥 ∈ ℂ𝑛 : (𝐴 − 𝜆1 𝐼)𝑘 𝑥 = 0 for some 𝑘 }.
(12.22)
Proof. Let 𝜆1 correspond to the matrix, say, 𝐴′1 in (12.18) and suppose (𝐴 − 𝜆1 𝐼)𝑘 𝑥 = 0 for some 𝑘. By setting 𝑥 = 𝑆𝑥′ and noting that the matrices 𝐴′𝑗 − 𝜆1 𝐼𝑛𝑗 from (12.18) are non-singular for 𝑗 ∕= 1 it follows ⎡
⎤ 𝑥′1 ⎢ 0 ⎥ ⎢ ⎥ 𝑥′ = ⎢ . ⎥ ⎣ .. ⎦
(12.23)
0 (0)
that is, 𝑥′ ∈ ℛ(𝑃1 ) or equivalently 𝑥 ∈ ℛ(𝑃1 ). Conversely, if 𝑥 ∈ ℛ(𝑃1 ) that is, (12.23) holds then ⎡
⎤ (𝐴′1 − 𝜆1 𝐼𝑛1 )𝑘 𝑥′1 ⎢ ⎥ 0 ⎢ ⎥ (𝐴 − 𝜆1 𝐼)𝑘 𝑥 = 𝑆(𝐴′ − 𝜆1 𝐼𝑛1 )𝑘 𝑥′ = 𝑆 ⎢ ⎥. .. ⎣ ⎦ . 0
12 Spectral Decomposition of a General 𝐽-Hermitian Matrix
99
Now the matrix 𝐴′1 has a single eigenvalue 𝜆1 of multiplicity 𝑛1 . Since we know that 𝜆1 is a pole of the resolvent (𝜆𝐼𝑛1 − 𝐴′1 )−1 the function (𝜆 − 𝜆1 )𝑘 (𝜆𝐼𝑛1 − 𝐴′1 )−1 will be regular at 𝜆 = 𝜆1 for some 𝑘. Thus, ∫ 1 (𝐴′1 − 𝜆1 𝐼𝑛1 )𝑘 = (𝜆 − 𝜆1 )𝑘 (𝜆𝐼𝑛1 − 𝐴′1 )−1 𝑑𝜆 = 0. 2𝜋𝑖 𝛤1 This implies (𝐴 − 𝜆1 𝐼)𝑘 𝑥 = 0. Q.E.D. The matrix
𝑁1 = 𝐴′1 − 𝜆1 𝐼𝑛1 ,
has the property that some power vanishes. Such matrices are called nilpotent. We now can make the spectral decomposition (12.18) more precise ⎡ ⎢ 𝐴′ = 𝑆 −1 𝐴𝑆 = ⎣
𝜆1 𝐼𝑛1 + 𝑁1
⎤ ..
⎥ ⎦
.
(12.24)
𝜆𝑝 𝐼𝑛𝑝 + 𝑁𝑝 with
𝜎(𝐴) = {𝜆1 , . . . , 𝜆𝑝 },
𝑁1 , . . . , 𝑁𝑝 nilpotent.
Obviously, the root space ℰ𝜆𝑘 contains the eigenspace; if the two are equal the eigenvalue is called semisimple or non-defective, otherwise it is called defective, the latter case is characterised by a non-vanishing nilpotent 𝑁𝑘 in (12.24). Similarly a matrix is called defective, if it has at least one defective eigenvalue. A defective matrix is characterised by the fact that its eigenvectors cannot form a basis of the whole space. The Jordan form of a nilpotent matrix will be of little interest in our considerations. If the matrix 𝐴 is real then the spectral sets 𝜎𝑘 can be chosen as symmetric with respect to the real axis: 𝜎𝑘 = 𝜎𝑘 . Then the curve 𝛤 , surrounding 𝜎𝑘 can also be chosen as 𝛤 = 𝛤 and we have ∫ ∫ 1 1 𝑃 𝜎𝑘 = − (𝜆𝐼 − 𝐴)−1 𝑑𝜆 = (𝜆𝐼 − 𝐴)−1 𝑑𝜆 = 𝑃𝜎𝑘 . 2𝜋𝑖 𝛤 2𝜋𝑖 𝛤 So, all subspaces ℛ(𝑃𝜎𝑘 ) are real and the matrix 𝑆 appearing in (12.18) can be chosen as real. Another case where this partition is convenient is if 𝐴 is 𝐽-Hermitian (even if complex) since 𝜎(𝐴) comes in complex conjugate pairs as well. Here we have (𝜆𝐼 − 𝐴)−∗ = 𝐽(𝜆𝐼 − 𝐴)−1 𝐽
100
12 Spectral Decomposition of a General 𝐽-Hermitian Matrix
and 𝑃𝜎∗𝑘 =
1 −2𝜋𝑖
∫ 𝛤
(𝜆𝐼 − 𝐴)−1 𝑑𝜆 =
1 2𝜋𝑖
∫ 𝛤
𝐽(𝜆𝐼 − 𝐴)−1 𝐽𝑑𝜆 = 𝐽𝑃𝜎𝑘 𝐽,
i.e. the projections 𝑃𝜎𝑘 are 𝐽-orthogonal, hence by Theorem 8.2 the matrix 𝑆 in (12.18) can be taken as 𝐽, 𝐽 ′ -unitary with ⎡ ⎢ 𝐽 ′ = 𝑆 ∗ 𝐽𝑆 = ⎣
𝐽1′
⎤ ..
⎥ ⎦.
.
(12.25)
𝐽𝑝′ In this case we call (12.18) a 𝐽, 𝐽 ′ unitary decomposition of a 𝐽-Hermitian matrix 𝐴. Achieving (12.24) with a 𝐽, 𝐽 ′ -unitary 𝑆 is possible if and only if all eigenvalues are real. Indeed, in this case 𝐴′ from (12.24) is 𝐽 ′ -Hermitian, hence each 𝜆𝑘 𝐼𝑛𝑘 + 𝑁𝑘 is 𝐽𝑘′ -Hermitian. Since its spectrum consists of a single point it must be real. Proposition 12.9 If the spectral part 𝜎𝑘 consists of 𝐽-definite eigenvalues (not necessarily of the same sign) then all these eigenvalues are non-defective and 𝜄± (𝐽𝑃𝜎𝑘 ) yields the number of the 𝐽-positive and 𝐽-negative eigenvalues in 𝜎𝑘 , respectively (counting multiplicities) and each root space is equal to the eigenspace. If all eigenvalues in 𝜎𝑘 are of the same type then in (12.25) we have 𝐽𝑘′ = ±𝐼𝑛𝑘 . Proof. The non-defectivity follows from Proposition 10.1. All other statements follow immediately from the identity (8.2) and Corollary 10.2 applied to the corresponding 𝐴′𝑘 from (12.18). Q.E.D. Exercise 12.10 Let 𝐴 be 𝐽-Hermitian and 𝜆𝑗 a 𝐽-definite eigenvalue. Then (i) 𝜆𝑗 is non-defective i.e. 𝑁𝑗 = 0 in (12.26) and (ii) 𝜆𝑗 is a simple pole of the resolvent (𝜆𝐼 − 𝐴)−1 . The 𝐽, 𝐽 ′ -unitary decomposition (12.18) can be further refined until each 𝐴′𝑗 has either a single real eigenvalue 𝜆𝑗 in which case it reads 𝐴′𝑗 = 𝜆𝑗 𝐼𝑛𝑗 + 𝑁𝑗′ , 𝑁𝑗′ nilpotent
(12.26)
(see Proposition 12.8) or its spectrum is {𝜆, 𝜆}, 𝜆 ∕= 𝜆. This is the ‘most refined’ decomposition which can be achieved with a diagonal matrix 𝐽 ′ , here also the real arithmetic is kept, if 𝐴 was real. If we allow for complex 𝐽, 𝐽 ′ -unitary transformations and non-diagonal 𝐽 ′ then a further blockdiagonalisation is possible. Proposition 12.11 Let 𝐴 be 𝐽-Hermitian of order 𝑛 and 𝜎(𝐴) = {𝜆, 𝜆}, 𝜆 ∕= 𝜆. Then 𝑛 is even and there exists a complex 𝐽, 𝐽 ′ -unitary matrix 𝑈 such that
12 Spectral Decomposition of a General 𝐽-Hermitian Matrix
𝑈
−1
101
[
] [ ] 𝜶 0 0𝐼 ′ ∗ 𝐴𝑈 = , 𝐽 = 𝑈 𝐽𝑈 = , 0 𝜶∗ 𝐼0
where 𝜶 = 𝜆𝐼 + 𝑁 and 𝑁 is a nilpotent matrix of order 𝑛/2. ˙ 𝜆 and dim(ℰ𝜆 ) = dim(ℰ𝜆 ) Proof. We know that in this case ℂ𝑛 = ℰ𝜆 +ℰ (Exercise 12.4). Now we need the following fact: Any root space corresponding to a non-real eigenvalue is 𝐽-neutral. Indeed, let (𝐴 − 𝜆𝐼)𝑘 𝑥 = 0, (𝐴 − 𝜇𝐼)𝑙 𝑦 = 0 for some 𝑘, 𝑙 ≥ 1. For 𝑘 = 𝑙 = 1 we have 0 = 𝑦 ∗ 𝐽𝐴𝑥 − 𝜆𝑦 ∗ 𝐽𝑥 = 𝑦 ∗ 𝐴∗ 𝐽𝑥 − 𝜆𝑥∗ 𝐽𝑥 = (2 Im 𝜆)𝑦 ∗ 𝐽𝑥
(12.27)
hence, 𝑦 ∗ 𝐽𝑥 = 0, which is already known from Theorem 9.1. For 𝑘 = 2, 𝑙 = 1, we set 𝑥 ˆ = (𝐴− 𝜆𝐼)𝑥, then 𝑥 ˆ ∈ ℰ𝜆 and (𝐴− 𝜆𝐼)ˆ 𝑥 = 0 and by applying (12.27) to 𝑥ˆ we have 0 = 𝑦∗𝐽 𝑥 ˆ = 𝑦 ∗ (𝐴 − 𝜆𝐼)𝑥 = (2 Im 𝜆)𝑦 ∗ 𝐽𝑥. The rest is induction. Now let the matrix 𝑋+ , 𝑋− carry the basis of ℰ𝜆 , ℰ𝜆 , respectively, in its columns. Then 𝑋 = [𝑋+ 𝑋− ] is non-singular and
[
0 𝐵 𝑋 𝐽𝑋 = 𝐵∗ 0 ∗
]
hence 𝐵 is square and non-singular. By using the polar decomposition 𝐵 = 𝑈 (𝐵 ∗ 𝐵)1/2 and
𝑌+ = 𝑋+ 𝑈 (𝐵 ∗ 𝐵)−1/2 𝑌− = 𝑋− (𝐵 ∗ 𝐵)−1/2
we see that 𝑌 = [ 𝑌+ 𝑌− ] satisfies 𝑌 ∗ 𝐽𝑌 = 𝐽 ′ = and also
[
0𝐼 𝐼0
]
𝐴𝑌+ = 𝑌+ 𝜶, 𝐴𝑌− = 𝑌− 𝜶−
i.e. 𝑌 −1 𝐴𝑌 =
[
] 𝜶 0 . 0 𝜶−
Since the last matrix is 𝐽 ′ -Hermitian by (6.1) we have 𝜶− = 𝜶∗ . Q.E.D.
102
12 Spectral Decomposition of a General 𝐽-Hermitian Matrix
Note that the block-diagonalisation in which each diagonal block corresponds to a single eigenvalue cannot be carried out in real arithmetic, if there are non-real eigenvalues hence the matrix 𝜶 cannot be real. Thus refined, we call (12.18) a complex spectral decomposition of a 𝐽Hermitian matrix 𝐴. If a non-real eigenvalue 𝜆𝑘 is semisimple then we can always choose 𝐴′𝑘 = ∣𝜆𝑘 ∣
[
] 0 𝐼𝑛𝑘 /2 , −𝐼𝑛𝑘 /2 0
𝐽𝑘′ =
[
] 𝐼𝑛𝑘 /2 0 . 0 −𝐼𝑛𝑘 /2
(12.28)
Exercise 12.12 Derive the formula (12.28). In view of ‘the arithmetic of pluses and minuses’ in Theorems 8.1, 8.2 and Proposition 12.11 the presence of 2𝑠 non-real spectral points (including multiplicities) ‘consumes’ 𝑠 pluses and 𝑠 minuses from the inertia of 𝐽. Consequently, 𝐴 has at least ∣𝜄+ (𝐽) − 𝜄− (𝐽)∣ = 𝑛 − 2𝜄+ (𝐴) real eigenvalues. Example 12.13 Consider the matrix ⎡
⎤ 6 −3 −8 𝐴 = ⎣ 3 −2 −4 ⎦ , 8 −4 −9 which is 𝐽-Hermitian with ⎡
⎤ 1 0 0 𝐽 = ⎣ 0 −1 0 ⎦ . 0 0 −1 The matrix
⎡
⎤ 3 −2 −2 𝑈 = ⎣ 2 −1 −2 ⎦ 2 −2 −1
is 𝐽-unitary and induces the block-diagonalisation ⎡
⎤ −2 1 0 𝑈 −1 𝐴𝑈 = ⎣ −1 −2 0 ⎦ 0 0 −1
12 Spectral Decomposition of a General 𝐽-Hermitian Matrix
whereas
103
√ √ √ ⎡ √ ⎤ 3√2/2 − 2𝑖 3 2/2 + √ √ √ 2𝑖 −2 ⎦ 𝑆 = ⎣ 2 − 2𝑖/2 2√+ 2𝑖/2 √ √ √ −2 2 − 2𝑖 2 + 2𝑖 −1
is 𝐽, 𝐽 ′ -unitary with
⎡
⎤ 01 0 𝐽′ = ⎣ 1 0 0 ⎦ 0 0 −1
and induces the block-diagonalisation (in fact, the full diagonalisation) ⎤ −2 + 𝑖 0 0 𝑆 −1 𝐴𝑆 = ⎣ 0 −2 − 𝑖 0 ⎦ . 0 0 −1 ⎡
There is no general information on the non-vanishing powers of the nilpotents appearing in the 𝐽, 𝐽 ′ -unitary decomposition of a 𝐽-Hermitian matrix. More can be said, if the matrix pair 𝐽𝐴, 𝐽 is semidefinite. Theorem 12.14 Let 𝐴 be 𝐽-Hermitian, 𝐽𝐴 − 𝜆𝐽 positive semidefinite and 𝐽𝐴 − 𝜆′ 𝐽 not positive or negative definite for any other real 𝜆′ . Then 𝜆 is an eigenvalue and in the decomposition (12.18) for 𝜆𝑝 = 𝜆 we have 𝐴′𝑝 = 𝜆𝑝 𝐼𝑛𝑝 + 𝑁𝑝 ,
𝑁𝑝2 = 0
and 𝑁𝑗 = 0 for 𝜆𝑗 ∕= 𝜆𝑝 . Proof. According to Theorem 10.6 𝜆 must be an eigenvalue of 𝐴 – otherwise 𝐽𝐴− 𝜆𝐽 would be positive definite and we would have a non void definiteness interval (points with 𝐽𝐴−𝜆𝐽 negative definite are obviously precluded). Also, with 𝐽 ′ from (12.25) the matrix 𝐽 ′ 𝐴′ −𝜆𝐽 ′ and therefore each block 𝐽𝑗′ 𝐴′𝑗 −𝜆𝐽𝑗′ is positive semidefinite and 𝜆 is the only spectral point of 𝐴′𝑗 . From the general theory 𝑁𝑗 = 𝐴′𝑗 − 𝜆𝐼𝑛𝑗 must be nilpotent i.e. 𝑁𝑗𝑟 = 0 for some 𝑟 ≤ 𝑛. At the same time 𝐽𝑗′ 𝑁𝑗 is positive semidefinite. Assuming without loss of generality that 𝑟 is odd and writing 𝑟 = 1 + 2𝑞 the equality 𝑁𝑗𝑟 = 0 implies 0 = 𝑥∗ 𝐽𝑁𝑗𝑟 𝑥 = (𝑁𝑗𝑞 𝑥)∗ 𝐽𝑁𝑗 𝑁𝑗𝑞 𝑥, which by the positive semidefiniteness of 𝐽𝑁𝑗 implies 𝑁𝑗𝑞 = 0. By induction this can be pushed down to 𝑁𝑗2 = 0. Q.E.D. In the theorem above the eigenvalue 𝜆 need not be defective, an example is 𝐴 = 𝐼. The developed spectral theory can be used to extend Theorem 10.5 to cases in which only some spectral parts of 𝐴 are 𝐽-definite. They, too, will remain such until they cross eigenvalues of different type. As in Theorem 10.5
104
12 Spectral Decomposition of a General 𝐽-Hermitian Matrix
we will have to ‘erect barriers’ in order to prevent the crossing, but now we have to keep out the complex eigenvalues as well. Theorem 12.15 Let 𝐴 be 𝐽-Hermitian and 𝜎0 ⊆ 𝜎(𝐴) consist of 𝐽-positive eigenvalues and let 𝛤 be any contour separating 𝜎0 from the rest of 𝜎(𝐴). Let ℐ ∋ 𝑡 → 𝐴(𝑡) be 𝐽-Hermitian-valued continuous function on a closed interval ℐ such that 𝐴 = 𝐴(𝑡0 ) for some 𝑡0 ∈ ℐ and such that 𝛤 ∩ 𝜎(𝐴(𝑡0 )) = ∅. Then the part of 𝜎(𝐴(𝑡)) within 𝛤 consists of 𝐽-positive eigenvalues whose number (with multiplicities) is constant over ℐ (and analogously in the 𝐽negative case). Proof. The key fact is the continuity of the total projection ∫ 1 𝑃 (𝑡) = (𝜆𝐼 − 𝐴(𝑡))−1 𝑑𝜆 2𝜋𝑖 𝛤 as a function of 𝑡, see Exercise 12.2. We can cover ℐ by a finite family of open intervals such that within each of them the quantity ∥𝑃 (𝑡′ ) − 𝑃 (𝑡)∥ is less than 1/2. Now use Theorem 8.4; it follows that all 𝑃 (𝑡) are 𝐽-unitarily similar. Then the number of the eigenvalues within 𝛤 , which equals Tr𝑃 (𝑡), is constant over ℐ and all 𝑃 (𝑡) are 𝐽-positive (Corollary 8.5). Q.E.D. Note that as in Theorem 10.5 the validity of the preceding theorem will persist, if we allow 𝛤 to move continuously in 𝑡. Example 12.16 We shall derive the spectral decomposition of the phasespace matrix corresponding to a modally damped system which is characterised by the property (2.24). Then according to Theorem 2.3 there is a (real) non-singular 𝛷 such that 𝛷𝑇 𝑀 𝛷 = 𝐼, 𝛷𝑇 𝐾𝛷 = 𝛺 2 = diag(𝜔12 , . . . , 𝜔𝑛2 ), 𝜔𝑖 > 0, 𝛷𝑇 𝐶𝛷 = 𝐷 = diag(𝑑11 , . . . , 𝑑𝑛𝑛 ), 𝑑𝑖𝑖 ≥ 0. The matrices 𝑈1 = 𝐿𝑇1 𝛷𝛺 −1 ,
𝑈2 = 𝐿𝑇2 𝛷𝛺 −1
are unitary and [ 𝐴= where
] 0 𝑈1 𝛺𝑈2−1 = 𝑈 𝐴′ 𝑈 −1 , 𝑈2 𝛺𝑈1−1 𝑈2 𝐷𝑈2−1 [
] 𝑈1 0 𝑈= , 0 𝑈2
[
0 𝛺 𝐴 = −𝛺 −𝐷 ′
]
and 𝑈 is jointly unitary. This 𝐴′ is essentially block-diagonal – up to a permutation. Indeed, define the permutation matrix 𝑉0 , given by 𝑉0 𝑒2𝑗−1 = 𝑒𝑗 , 𝑉0 𝑒2𝑗 = 𝑒𝑗+𝑛 (this is the so called perfect shuffling). For 𝑛 = 3 we have
12.1 Condition Numbers
105
⎡
100 ⎢0 0 1 ⎢ ⎢ ⎢0 0 0 𝑉0 = ⎢ ⎢0 1 0 ⎢ ⎣0 0 0 000
00 00 01 00 10 00
⎤ 0 0⎥ ⎥ ⎥ 0⎥ ⎥. 0⎥ ⎥ 0⎦ 1
The matrix 𝑉0 is unitary and 𝐽, 𝐽 ′ -unitary with 𝐽 ′ = diag(𝑗, . . . , 𝑗), and we have
[ 𝑗=
1 0 0 −1
]
𝐴′′ = 𝑉0−1 𝐴′ 𝑉0 = diag(𝐴′′1 , . . . , 𝐴′′𝑛 ), [ ] 0 𝜔𝑖 𝐴′′𝑖 = . −𝜔𝑖 −𝑑𝑖𝑖
(12.29)
Altogether, 𝑉 = 𝑈 𝑉0 is 𝐽, 𝐽 ′ -unitary (and unitary) and we have the real spectral decomposition 𝑉 −1 𝐴𝑉 = 𝐴′′ . Exercise 12.17 Use perfect shuffling and a convenient choice of 𝐿1 , 𝐿2 in (2.16) so as to make the phase-space matrix for our model example in Chap. 1 as narrowly banded as possible. Exercise 12.18 Derive the analog of (12.29) for the phase-space matrix (4.19).
12.1
Condition Numbers
We now turn back to the question of the condition number in blockdiagonalising a 𝐽-Hermitian matrix. More precisely, we ask: are the 𝐽, 𝐽 ′ -unitaries better than others, if the block-diagonalised matrix was 𝐽-Hermitian? The answer is positive. This makes 𝐽, 𝐽 ′ unitaries even more attractive in numerical applications. The rest of this chapter will be devoted to this question. Most interesting is the case in which in (12.18), (12.19) the spectra of 𝐴′𝑖 are disjoint and symmetric with respect to the real axis, i.e. 𝜎(𝐴) is partitioned as 𝜎𝑘 = 𝜎(𝐴′𝑘 ),
𝜎(𝐴) = 𝜎1 ∪ ⋅ ⋅ ⋅ ∪ 𝜎𝑝 , 𝜎 𝑗 = 𝜎𝑗 , 𝑗 = 1, . . . , 𝑝, 𝜎𝑖 ∩ 𝜎𝑗 = ∅ if 𝑖 ∕= 𝑗.
106
12 Spectral Decomposition of a General 𝐽-Hermitian Matrix
Then the relation (8.6) is equivalent to the block-diagonalisation (12.18) and it is sufficient to study the transformation (8.6). Theorem 12.19 Let 𝑃1 , . . . 𝑃𝑝 be any 𝐽-orthogonal decomposition of the identity and 𝑆 non-singular with (0)
𝑆 −1 𝑃𝑗 𝑆 = 𝑃𝑗 , 𝑗 = 1, . . . , 𝑝 (0)
with 𝑃𝑗
(12.30)
from (12.19). Then there is a 𝐽, 𝐽 ′ -unitary 𝑈 such that ⎡
⎢ (0) 𝑈 −1 𝑃𝑗 𝑈 = 𝑃𝑗 , 𝐽 ′ = ⎣
𝐽1′
⎤ ..
. 𝐽𝑝′
and
⎥ ′ , −𝐼𝑛− ) ⎦ , 𝐽𝑘 = diag(𝐼𝑛+ 𝑘 𝑘
𝜅(𝑈 ) ≤ 𝜅(𝑆).
The same inequality holds for the condition number measured in the Euclidian norm. Proof. Any 𝑆 satisfying (12.30) will be called block-diagonaliser. By representing 𝑆 as ] [ 𝑆 = 𝑆1 ⋅ ⋅ ⋅ 𝑆𝑝 (the partition as in (12.18)) we can write (12.30) as 𝑆𝑘 𝑃𝑗 = 𝛿𝑘𝑗 𝑃𝑘 . Then 𝑆 ∗ 𝐽𝑆 is block-diagonal: 𝑆𝑗∗ 𝐽𝑆𝑘 = 𝑆𝑗∗ 𝐽𝑃𝑘 𝑆𝑘 = 𝑆𝑗∗ 𝑃𝑘∗ 𝐽𝑆𝑘 = (𝑃𝑘 𝑆𝑗 )∗ 𝐽𝑆𝑘 = 𝛿𝑘𝑗 𝑆𝑗∗ 𝐽𝑆𝑘 . Hence each matrix 𝑆𝑘∗ 𝐽𝑆𝑘 is Hermitian and non-singular; its eigenvalue decomposition can be written as 𝑆𝑘∗ 𝐽𝑆𝑘 = 𝑊𝑘 𝛬𝑘 𝑊𝑘∗ = 𝑊𝑘 ∣𝛬𝑘 ∣1/2 𝐽𝑘 ∣𝛬𝑘 ∣1/2 𝑊𝑘∗ ,
𝑘 = 1, . . . , 𝑝,
where 𝛬𝑘 is diagonal and non-singular, 𝑊𝑘 is unitary and 𝐽𝑘 = sign(𝛬𝑘 ). By setting 𝛬 = diag(𝛬1 , . . . , 𝛬𝑝 ), the matrix
𝐽 ′ = diag(𝐽1 , . . . , 𝐽𝑝 ), 𝑈 = 𝑆𝑊 ∣𝛬∣−1/2
𝑊 = diag(𝑊1 , . . . , 𝑊𝑝 )
12.1 Condition Numbers
107
is immediately seen to be 𝐽, 𝐽 ′ -unitary. Now, since 𝛬 and 𝐽 ′ commute, 𝜅(𝑆) = ∥𝑈 ∣𝛬∣1/2 𝑊 ∗ ∥∥𝑊 ∣𝛬∣−1/2 𝑈 −1 ∥ = ∥𝑈 ∣𝛬∣1/2 ∥∥𝐽 ′ ∣𝛬∣−1/2 𝑈 ∗ 𝐽∥ = ∥𝑈 ∣𝛬∣1/2 ∥∥𝛬∣−1/2 𝑈 ∗ ∥ ≥ ∥𝑈 𝑈 ∗ ∥ = 𝜅(𝑈 ). The proof for the Euclidian norm also uses the property 𝜅𝐸 (𝐹 ) = ∥𝐹 ∥𝐸 ∥𝐹 −1 ∥𝐸 = ∥𝐹 ∥𝐸 ∥𝐽1 𝐹 ∗ 𝐽∥𝐸 = ∥𝐹 ∥𝐸 ∥𝐹 ∗ ∥𝐸 = ∥𝐹 ∥2𝐸 , (12.31) valid for any 𝐽, 𝐽 ′ unitary 𝐹 . Thus, 𝜅𝐸 (𝑆)2 = Tr(𝑆 ∗ 𝑆)Tr(𝑆 −1 𝑆 −∗ ) = Tr(𝐹 ∗ 𝐹 ∣𝛬∣)Tr(∣𝛬∣−1 𝐹 ∗ 𝐹 ) =
( 𝑛 ∑
∣𝜆𝑘 ∣Tr(𝐹 ∗ 𝐹 )𝑘𝑘
)( 𝑛 ∑
𝑘
due to the inequality
) Tr(𝐹 ∗ 𝐹 )𝑘𝑘 /∣𝜆𝑘 ∣
≥ Tr(𝐹 ∗ 𝐹 )2
𝑘
∑
𝑝𝑘 𝜙𝑘
𝑘
∑
𝜙𝑘 /𝑝𝑘 ≥
∑
𝑘
𝜙2𝑘
𝑘
valid for any positive 𝑝𝑘 and non-negative 𝜙𝑘 . Q.E.D. According to Theorem 12.19 and Proposition 10.17 if the spectrum is definite then the best condition number of a block-diagonalising matrix is achieved, by any 𝐽, 𝐽 ′ -unitary among them. If the spectrum is not definite or if the spectral partition contains spectral subsets consisting of eigenvalues of various types (be these eigenvalues definite or not3 ) then the mere fact that a block-diagonalising matrix is 𝐽, 𝐽 ′ -unitary does not guarantee that the condition number is the best possible. It can, indeed, be arbitrarily large. But we know that a best-conditioned block-diagonaliser should be sought among the 𝐽, 𝐽 ′ -unitaries. The following theorem contains a construction of an optimal block-diagonaliser. Theorem 12.20 Let 𝑃1 , . . . , 𝑃𝑝 ∈ 𝛯 𝑛,𝑛 , 𝐽 ′ be as in Theorem 12.19. Then there is a 𝐽, 𝐽 ′ -unitary block-diagonaliser 𝐹 ∈ 𝛯 𝑛,𝑛 with ˆ 𝜅𝐸 (𝐹 ) ≤ 𝜅𝐸 (𝑆) ˆ for any block-diagonaliser 𝑆. Proof. We construct 𝐹 from an arbitrary given block-diagonaliser 𝑆 as [ ] follows. By setting 𝑆 = 𝑆1 ⋅ ⋅ ⋅ 𝑆𝑝 and using the properties established 3 Such
coarser partition may appear necessary, if a more refined partition with 𝐽-definite spectral sets would yield highly conditioned block-diagonaliser.
108
12 Spectral Decomposition of a General 𝐽-Hermitian Matrix
in the proof of Theorem 12.19 we may perform the generalised eigenvalue decomposition 𝛹𝑘∗ 𝑆𝑘∗ 𝑆𝑘 𝛹𝑘 = diag(𝛼𝑘1 , . . . , 𝛼𝑘𝑛+ , 𝛽1𝑘 , . . . , 𝛽𝑛𝑘− ),
(12.32)
𝛹𝑘∗ 𝑆𝑘∗ 𝐽𝑆𝑘 𝛹𝑘 = 𝐽𝑘′ = diag(𝐼𝑛+ , −𝐼𝑛− ),
(12.33)
𝑘
𝑘
𝑘
𝑘
𝑘 = 1, . . . , 𝑝. An optimal block-diagonaliser is given by ] [ 𝐹 = 𝐹1 ⋅ ⋅ ⋅ 𝐹𝑝 , 𝐹𝑘 = 𝑆𝑘 𝛹𝑘 .
(12.34)
It is clearly 𝐽, 𝐽 ′ -unitary by construction. To prove the optimality first note that the numbers 𝑛𝑘± and 𝛼𝑘𝑗 , 𝛽𝑗𝑘 do not depend on the block-diagonaliser 𝑆. Indeed, any block-diagonaliser 𝑆ˆ is given by ] [ 𝑆ˆ = 𝑆1 𝛤1 ⋅ ⋅ ⋅ 𝑆𝑝 𝛤𝑝 ,
𝛤1 , . . . , 𝛤𝑝 non-singular.
By the Sylvester theorem 𝑛𝑘± is given by the inertia of 𝑆𝑘∗ 𝐽𝑆𝑘 which is the same as the one of 𝑆ˆ𝑘∗ 𝐽 𝑆ˆ𝑘 = 𝛤𝑘∗ 𝑆𝑘∗ 𝐽𝑆𝑘 𝛤𝑘 . Furthermore, the numbers 𝛼𝑘𝑗 , −𝛽𝑗𝑘 are the generalised eigenvalues of the matrix pair 𝑆𝑘∗ 𝑆𝑘 , 𝑆𝑘∗ 𝐽𝑆𝑘 and they do not change with the transition to the congruent pair 𝛤𝑘∗ 𝑆𝑘∗ 𝑆𝑘 𝛤𝑘 , 𝛤𝑘∗ 𝑆𝑘∗ 𝐽𝑆𝑘 𝛤𝑘 . In particular, the sums 𝑡𝑘0 = 𝛼𝑘1 + ⋅ ⋅ ⋅ + 𝛼𝑘𝑛+ + 𝛽1𝑘 + ⋅ ⋅ ⋅ + 𝛽𝑛𝑘− ,
𝑡0 =
𝑝 ∑
𝑡𝑘0
𝑘=1
are independent of 𝑆. By Theorem 12.19 there is a 𝐽, 𝐽 ′ -unitary blockdiagonaliser 𝑈 with 𝜅𝐸 (𝑈 ) ≤ 𝜅𝐸 (𝑆). (12.35) If we now do the construction (12.32)–(12.34) with 𝑆 replaced by 𝑈 then (12.33) just means that 𝛹𝑘 is 𝐽𝑘′ -unitary. From Theorem 10.16 and from the above mentioned invariance property of 𝑡𝑘0 it follows 𝑡𝑘0 = Tr(𝐹𝑘∗ 𝐹𝑘 ) ≤ Tr(𝑈𝑘∗ 𝑈𝑘 ) and, by summing over 𝑘, 𝑡0 = Tr(𝐹 ∗ 𝐹 ) ≤ Tr(𝑈 ∗ 𝑈 ). By (12.31) this is rewritten as 𝑡0 = 𝜅𝐸 (𝐹 )2 ≤ 𝜅𝐸 (𝑈 )2 . This with (12.35) and the fact that 𝑆 was arbitrary gives the statement. Q.E.D.
12.1 Condition Numbers
109
The numbers 𝛼𝑘𝑖 , 𝛽𝑗𝑘 appearing in the proof of the previous theorem have a deep geometrical interpretation which is given in the following theorem. Theorem 12.21 Let 𝐹 be any optimal block-diagonaliser from Theorem 12.20. Then the numbers 𝛼𝑘𝑖 , 𝛽𝑗𝑘 from (12.32) are the non-vanishing singular values of the projection 𝑃𝑘 . Moreover, 𝜅𝐸 (𝐹 ) = ∥𝐹 ∥2𝐸 =
𝑝 ∑ 𝑘=1
⎛ + ⎞ 𝑛𝑘 𝑛− 𝑝 𝑘 ∑ ∑ ∑ √ ⎝ 𝛼𝑘𝑖 + 𝛽𝑖𝑘 ⎠ = Tr 𝑃𝑘 𝑃 ∗ . 𝑖=1
𝑖=1
𝑘=1
𝑘
(12.36)
Proof. We revisit the proof of Theorem 12.20. For any fixed 𝑘 we set 𝜶 = diag(𝛼𝑘1 , . . . , 𝛽𝑛𝑘− ). Since the numbers 𝛼𝑘1 , . . . , 𝛽𝑛𝑘− do not depend on the block 𝑘
𝑘
diagonaliser 𝑆 we may take the latter as 𝐽, 𝐽 ′ -unitary: [ ] 𝑆 = 𝑈 = 𝑈1 ⋅ ⋅ ⋅ 𝑈𝑝 ,
𝑈𝑗∗ 𝐽𝑈𝑘 = 𝛿𝑘𝑗 𝐽𝑘′
𝑃𝑘 = 𝑈𝑘 𝐽𝑈𝑘∗ 𝐽𝑘′
Now (12.32) and (12.33) read 𝛹𝑘∗ 𝑈𝑘∗ 𝑈𝑘 𝛹𝑘 = 𝜶,
𝛹𝑘∗ 𝐽𝑘′ 𝛹𝑘 = 𝐽𝑘′
and we have 𝐹𝑘 = 𝑈𝑘 𝛹𝑘 . It is immediately verified that the matrix 𝑉 = 𝑈𝑘 𝛹𝑘 𝜶−1/2 is an isometry. Moreover, its columns are the eigenvectors to the non-vanishing eigenvalues of 𝑃𝑘 𝑃𝑘∗ . To prove this note the identity 𝑃𝑘 𝑃𝑘∗ = (𝑈𝑘 𝐽𝑘′ 𝑈𝑘∗ )2 . On the other hand, using 𝛹𝑘 𝐽𝑘′ 𝛹𝑘∗ = 𝐽𝑘 gives 𝑈𝑘 𝐽𝑘′ 𝑈𝑘∗ 𝑉 = 𝑈𝑘 𝛹𝑘 𝐽𝑘′ 𝛹𝑘∗ 𝑈𝑘∗ 𝑈𝑘 𝛹𝑘 𝜶−1/2 = 𝑈𝑘 𝛹𝑘 𝐽𝑘′ 𝜶1/2 and, since both 𝜶 and 𝐽𝑘′ are diagonal, 𝑈𝑘 𝐽𝑘′ 𝑈𝑘∗ 𝑉 = 𝑉 𝐽𝑘′ 𝜶, so the diagonal of 𝐽𝑘′ 𝜶 consists of all non-vanishing eigenvalues of the Hermitian matrix 𝑈𝑘 𝐽𝑘′ 𝑈𝑘∗ . Hence the diagonal of 𝜶 consists of the nonvanishing singular values of 𝑃𝑘 . Then also +
∥𝐹𝑘 ∥2𝐸 = Tr𝜶 =
𝑛𝑘 ∑ 𝑖=1
so (12.36) holds as well. Q.E.D.
−
𝛼𝑘𝑖 +
𝑛𝑘 ∑ 𝑖=1
√ 𝛽𝑖𝑘 = Tr 𝑃𝑘 𝑃𝑘∗ ,
110
12 Spectral Decomposition of a General 𝐽-Hermitian Matrix
√ The value Tr 𝑃𝑘 𝑃𝑘∗ (this is, in fact, a matrix norm) may be seen as a condition number of the projection 𝑃𝑘 . Indeed, we have 𝑛𝑘 = Tr𝑃𝑘 ≤ Tr
√ 𝑃𝑘 𝑃𝑘∗ ,
(12.37)
where equality is attained, if and only if the projection 𝑃𝑘 is orthogonal in the ordinary sense. To prove (12.37) we write the singular value decomposition of 𝑃𝑘 as 𝑃𝑘 = 𝑈 𝜶𝑉 ∗ , where 𝑈, 𝑉 are isometries of type 𝑛 × 𝑛𝑘 . The identity 𝑃𝑘2 = 𝑃𝑘 means (note that 𝜶 is diagonal positive definite) 𝜶𝑉 ∗ 𝑈 = 𝐼𝑛𝑘 . Now
𝑛𝑘 = Tr𝑃𝑘 = Tr(𝑈 𝜶𝑉 ∗ ) = (𝑉 𝜶1/2 , 𝑈 𝜶1/2 )𝐸 ,
where
(𝑇, 𝑆)𝐸 = Tr(𝑆 ∗ 𝑇 )
denotes the Euclidian scalar product on matrices. Using the Cauchy-Schwartz inequality and the fact that 𝑈, 𝑉 are isometries we obtain √ 𝑛𝑘 = Tr𝑃𝑘 ≤ ∥𝑉 𝜶1/2 ∥𝐸 ∥𝑈 𝜶1/2 ∥𝐸 ≤ ∥𝜶1/2 ∥2𝐸 = Tr𝜶 = Tr 𝑃𝑘 𝑃𝑘∗ . If this inequality turns to an equality then the matrices 𝑉 𝜶1/2 , 𝑈 𝜶1/2 , and therefore 𝑉, 𝑈 must be proportional, which implies that 𝑃𝑘 is Hermitian and therefore orthogonal in the standard sense. If this is true for all 𝑘 then by (12.36) the block-diagonaliser 𝐹 is jointly unitary. There are simple damped systems in which no reduction like (12.18) is possible. One of them was produced as the critical damping case in Example 9.3. Another is given by the following Example 12.22 Take the system in (1.2)–(1.4) with 𝑛 = 2 and 𝑚1 = 1, 𝑚2 =
25 , 4
𝑘1 = 0, 𝑘2 = 5, 𝑘3 = [ 𝐶=
] 40 . 00
5 , 4
12.1 Condition Numbers
111
Tedious, but straightforward calculation (take 𝐿1 , 𝐿2 as the Cholesky factors) shows that in this case the matrix (3.3) looks like √ √ ⎤ 0 0 5 −2/√ 5 ⎢ 0 0 0 1/ 5 ⎥ ⎥ = −𝐼 + 𝑁, √ 𝐴=⎢ ⎦ ⎣− 5 0 −4 0 √ √ 2/ 5 −1/ 5 0 0 ⎡
(12.38)
where 𝑁 3 ∕= 0, 𝑁 4 = 0. So, in (12.18) we have 𝑝 = 1 and no blockdiagonalisation is possible.4 Analogous damped systems can be constructed in any dimension. Exercise 12.23 Prove (12.38). Exercise 12.24 Modify the parameters 𝑚2 , 𝑘1 , 𝑘2 , 𝑘3 , 𝑐 in Example 12.22 in order to obtain an 𝐴 with a simple real spectrum (you may make numerical experiments). Exercise 12.25 Find a permutation 𝑃 such that 𝑃 𝑇 𝐴𝑃 (𝐴 from (12.38)) is tridiagonal.
4 This
intuitive fact is proved by constructing the Jordan form of the matrix 𝑁 which we omit here.
Chapter 13
The Matrix Exponential
The matrix exponential, its properties and computation play a vital role in a study of damped systems. We here derive some general properties, and then touch a central concern of oscillations: the exponential decay which is crucial in establishing stability of vibrating structures. The decomposition (12.18) can be used to simplify the computation of the matrix exponential needed in the solution (3.4) of the system (3.2). We have ⎡ ′ ⎢ 𝑒𝐴𝑡 = 𝑆𝑒𝐴 𝑡 𝑆 −1 = 𝑆 ⎣
⎤
′
𝑒 𝐴1 𝑡
..
⎥ −1 ⎦𝑆 ,
. 𝑒
(13.1)
𝐴′𝑝 𝑡
with a 𝐽, 𝐽 ′ -unitary 𝑆 and 𝐽 ′ from (12.25). As was said a ‘most refined’ decomposition (13.1) is obtained, if 𝜎(𝐴′𝑗 ) = 𝜆𝑗 is real, with 𝐴′𝑗 = 𝜆𝑗 𝐼𝑛𝑗 + 𝑁𝑗 ,
𝑁𝑗
nilpotent
or 𝜎(𝐴′𝑗 ) = {𝜆𝑗 , 𝜆𝑗 } non-real with 𝐽𝑗′ and
[
] 0 𝐼𝑛𝑗 /2 = , 𝐼𝑛𝑗 /2 0
𝜶𝑗 = 𝜆𝑗 𝐼𝑛𝑗 + 𝑀𝑗 ,
[
]
𝐴′𝑗
𝜶𝑗 0 = 0 𝜶∗𝑗
𝑀𝑗
nilpotent.
′
Thus, the computation of 𝑒𝐴 𝑡 reduces to exponentiating 𝑒(𝜆𝐼+𝑁 )𝑡 with 𝑁 nilpotent: 𝑟 ∑ 𝑡𝑘 𝑘 𝑒(𝜆𝐼+𝑁)𝑡 = 𝑒𝜆𝑡 𝑒𝑁 𝑡 = 𝑒𝜆𝑡 𝑁 , (13.2) 𝑘! 𝑘=0
K. Veseli´ c, Damped Oscillations of Linear Systems, Lecture Notes in Mathematics 2023, DOI 10.1007/978-3-642-21335-9 13, © Springer-Verlag Berlin Heidelberg 2011
113
114
13 The Matrix Exponential
where 𝑁 𝑟 is the highest non-vanishing power of 𝑁 . (In fact, the formulae (13.1), (13.2) hold for any matrix 𝐴, where 𝑆 has no 𝐽-unitarity properties, it is simply non-singular.) Example 13.1 We compute the matrix exponential to the matrix 𝐴 from (9.5), Example 9.3. First consider 𝐷 = 𝑑2 − 4𝜔 2 > 0 and use (9.7)–(9.9) to obtain ( √ [ ]) √ sinh( 𝐷𝑡/2) 𝑑 2𝜔 𝐴𝑡 −𝑑𝑡/2 √ 𝑒 =𝑒 cosh( 𝐷𝑡/2)𝐼 + . (13.3) −2𝜔 −𝑑 𝐷 A similar formula holds for 𝐷 < 0. This latter formula can be obtained from (13.3) by taking into account that 𝑒𝐴𝑡 is an entire function in each of the variables 𝜔, 𝑑, 𝑡. We use the identities √ √ cosh( 𝐷𝑡/2) = cos( −𝐷𝑡/2),
√ √ sin( −𝐷𝑡/2) sinh( 𝐷𝑡/2) √ √ = −𝐷 𝐷
thus obtaining 𝑒
𝐴𝑡
=𝑒
−𝑑𝑡/2
√ ( [ ]) √ sin( −𝐷𝑡/2) 𝑑 2𝜔 √ cos( −𝐷𝑡/2)𝐼 + , −2𝜔 −𝑑 −𝐷
(13.4)
which is more convenient for 𝐷 < 0. For 𝐷 = 0 using the L’Hospital rule in either (13.3) or (13.4) we obtain ( [ ]) 1 1 𝑒𝐴𝑡 = 𝑒−𝑑𝑡/2 𝐼 + 𝜔𝑡 . −1 −1 It does not seem to be easy to give an exponential bound for the norm of (13.3) and (13.4) which would be both simple and tight. By introducing the functions sinc(𝜏 ) = sin(𝜏 )/𝜏, sinhc(𝜏 ) = sinh(𝜏 )/𝜏 we can express both exponentials by the ‘special matrix function’ 𝑒2(𝜏, 𝜃) given as ([ ] ) 0 1 𝑒2(𝜏, 𝜃) = exp 𝜏 −1 −2𝜃 ⎧ ( [ ]) √ √ 𝜃 1 −𝜃𝜏 2 2 𝑒 cos( 1 − 𝜃 𝜏 )𝐼 + 𝜏 sinc( 1 − 𝜃 𝜏 ) 𝜃1 −1 −𝜃 ( [ ]) 1 1 −𝜏 𝐼 +𝜏 𝜃=1 ⎩𝑒 −1 −1
13 The Matrix Exponential
115
1.2
0
1
0.8
0.6 0.2 2.3
0.4
1.5 0.2
0
0.5 1 0
1
2
3
4
5
6
Fig. 13.1 Norm decay of 𝑒2
defined for any real 𝜏 and any non-negative real 𝜃. So, we have 𝑒𝐴𝑡 = 𝑒2(𝜔𝑡,
𝑑 ). 2𝜔
On Fig. 13.1 we have plotted the function 𝜏 → ∥𝑒2(𝜏, 𝜃)∥ for different values of 𝜃 shown on each graph. We see that different viscosities may produce locally erratic behaviour of ∥𝑒2(𝜏, 𝜃)∥ within the existing bounds for it. An ‘overall best behaviour’ is apparent at 𝜃 = 1 (critical damping, bold line). The 2×2 case studied in Example 13.1 is unexpectedly annoying: the formulae (13.3) or (13.4) do not allow a straightforward and simple estimate of the exponential decay of the matrix exponential. What can be done is to compute max 𝑒𝜈𝑡 ∥𝑒2(𝜔𝑡, 𝜃)∥ = max 𝑒𝜈𝜏 /𝜔 ∥𝑒2(𝜏, 𝜃)∥ = 𝑓 2(𝜅, 𝜃) 𝑡≥0
with
𝜏 ≥0
𝑓 2(𝜅, 𝜃) = max 𝑒𝜅𝜏 ∥𝑒2(𝜏, 𝜃)∥. 𝜏 ≥0
The function 𝑓 2 can be numerically tabulated and we have 𝜈 ∥𝑒2(𝜔𝑡)∥ ≤ 𝑒−𝜈𝑡 𝑓 2( , 𝜃). 𝜔
116
13 The Matrix Exponential
Remark 13.2 The function 𝑓 2 is defined whenever 𝜃 > 0 and 0 < 𝜅 = 𝜅0 , where √ { 𝜃 − 𝜃2 − 1, 𝜃 > 1 𝜅0 = 𝜃, 𝜃 0 there is a 𝐶𝜖 > 0 such that ∥𝑒𝐴𝑡 ∥ ≤ 𝐶𝜖 𝑒(𝜆𝑎 +𝜖)𝑡 , (13.7) where naturally one would be interested to find a best (that is, the smallest) 𝐶𝜖 for a given 𝜖 > 0. If ∥𝑒𝐴𝑡 ∥ ≤ 𝐹 𝑒−𝜈𝑡 (13.8) holds with some 𝜈 > 0, 𝐹 ≥ 1 and all 𝑡 ≥ 0 then we say that the matrix 𝐴 is asymptotically stable and 𝑒𝐴𝑡 exponentially decaying. The following theorem is fundamental. Theorem 13.6 For a general square matrix 𝐴 the following are equivalent: (i) (ii) (iii) (iv)
𝐴 is asymptotically stable. Re 𝜎(𝐴) < 0. 𝑒𝐴𝑡 → 0, 𝑡 → ∞. For some (and then for each) positive definite 𝐵 there is a unique positive definite 𝑋 such that 𝐴∗ 𝑋 + 𝑋𝐴 = −𝐵.
(13.9)
(v) ∥𝑒𝐴𝑡 ∥ < 1, for some 𝑡 = 𝑡1 > 0. In this case we have
∥𝑒𝐴𝑡 ∥ ≤ 𝜅(𝑋)𝑒−𝑡/𝛽
with 𝛽 = max 𝑦
(13.10)
𝑦 ∗ 𝑋𝑦 = ∥𝐵 −1/2 𝑋𝐵 −1/2 ∥. 𝑦 ∗ 𝐵𝑦
Proof. The equivalences (i) ⇐⇒ (ii) ⇐⇒ (iii) obviously follow from (13.1) and (13.2). Equation (13.9) is known as the Lyapunov equation and it can be understood as a linear system of 𝑛2 equations with as many unknowns. Thus, let (i) hold. Then the integral ∫ 𝑋=
1 This
0
∞
∗
𝑒𝐴 𝑡 𝐵𝑒𝐴𝑡 𝑑𝑡,
(13.11)
difficulty is somewhat alleviated, if the block-diagonalisation went up to diagonal blocks of dimension one or two only; then we may use the formulae derived in Example 13.1.
118
13 The Matrix Exponential
(which converges by (13.8)) represents a positive definite matrix which solves (13.9). Indeed, 𝑋 is obviously Hermitian and 𝑦 ∗ 𝑋𝑦 =
∫
∞
0
∗
𝑦 ∗ 𝑒𝐴 𝑡 𝐵𝑒𝐴𝑡 𝑦 𝑑𝑡 ≥ 0,
where the equality sign is attained, if and only if 𝐵𝑒𝐴𝑡 𝑦 ≡ 0, which implies 𝑦 = 0. The matrix 𝑋 solves the Lyapunov equation. Indeed, by partial integration we obtain ∗
∫
𝐴 𝑋=
∞ ∗ 𝑑 𝐴∗ 𝑡 𝐴𝑡 𝑒 𝐵𝑒 𝑑𝑡 = 𝑒𝐴 𝑡 𝐵𝑒𝐴𝑡 0 𝑑𝑡
∞ 0
∫ −
∞
0
∗
𝑒𝐴 𝑡 𝐵
𝑑 𝐴𝑡 𝑒 𝑑𝑡 = −𝐵 − 𝑋𝐴. 𝑑𝑡
The unique solvability of the Lyapunov equation under the condition (ii) is well-known; indeed the more general Sylvester equation 𝐶𝑋 + 𝑋𝐴 = −𝐵 is uniquely solvable, if and only if 𝜎(𝐶) ∩ 𝜎(−𝐴) = ∅ and this is guaranteed, if 𝐶 ∗ = 𝐴 is asymptotically stable. Thus (i) implies (iv). Now, assume that a positive definite 𝑋 solves (13.9). Then ∗ 𝑑 ( ∗ 𝐴∗ 𝑡 𝐴𝑡 ) 𝑦 𝑒 𝑋𝑒 𝑦 = 𝑦 ∗ 𝑒𝐴 𝑡 (𝐴∗ 𝑋 + 𝑋𝐴)𝑒𝐴𝑡 𝑦 𝑑𝑡 ∗
∗
= −𝑦 ∗ 𝑒𝐴 𝑡 𝐵𝑒𝐴𝑡 𝑦 ≤ −𝑦 ∗ 𝑒𝐴 𝑡 𝑋𝑒𝐴𝑡𝑦/𝛽 or
𝑑 ( ∗ 𝐴∗ 𝑡 𝐴𝑡 ) 1 ln 𝑦 𝑒 𝑋𝑒 𝑦 ≤ − . 𝑑𝑡 𝛽
We integrate this from zero to 𝑠 and take into account that 𝑒𝐴 ∗
𝑦 ∗ 𝑒𝐴 𝑠 𝑋𝑒𝐴𝑠 𝑦 ≤ 𝑦 ∗ 𝑋𝑦𝑒−𝑠/𝛽 which implies Hence
∗
𝑦 ∗ 𝑒𝐴 𝑠 𝑒𝐴𝑠 𝑦 ≤ 𝑦 ∗ 𝑋𝑦∥𝑋 −1∥𝑒−𝑠/𝛽 . ∥𝑒𝐴𝑠 ∥2 ≤ ∥𝑋∥∥𝑋 −1∥𝑒−𝑠/𝛽
∗
0
= 𝐼:
13 The Matrix Exponential
119
and (13.10) follows. This proves (iv) ⇒ (iii). Finally, (iii) ⇒ (v) is obvious; conversely, any 𝑡 can be written as 𝑡 = 𝑘𝑡1 + 𝜏 for some 𝑘 = 0, 1, . . . and some 0 ≤ 𝜏 ≤ 𝑡1 . Thus, ∥𝑒𝐴𝑡 ∥ = ∥(𝑒𝐴𝑡1 )𝑘 𝑒𝐴𝜏 ∥ ≤ ∥𝑒𝐴𝑡1 ∥𝑘 max ∥𝑒𝐴𝜏 ∥ → 0, 0≤𝜏 ≤𝑡1
𝑡 → ∞.
Thus, (v) ⇒ (iii). Q.E.D. Exercise 13.7 Suppose that 𝐴 is dissipative and 𝛼 = ∥𝑒𝐴𝑡1 ∥ < 1, for some 𝑡1 > 0. Show that (13.8) holds with 𝐹 = 1/𝛼,
𝜈=−
ln 𝛼 . 𝑡1
Corollary 13.8 If 𝐴 is asymptotically stable then 𝑋 from (13.11) solves (13.9) for any matrix 𝐵. Exercise 13.9 Show that for an asymptotically stable 𝐴 the matrix 𝑋 from (13.11) is the only solution of the Lyapunov equation (13.9) for an arbitrary 𝐵. Hint: supposing 𝐴∗ 𝑌 + 𝑌 𝐴 = 0 compute the integral ∫ ∞ ∗ 𝑒𝐴 𝑡 (𝐴∗ 𝑌 + 𝑌 𝐴)𝑒𝐴𝑡 𝑑𝑡. 0
Exercise 13.10 Show that already the positive semidefiniteness of 𝐵 almost always (i.e. up to rare special cases) implies the positive definiteness of 𝑋 in (13.9). Which are the exceptions? The matrix 𝑋 from (13.9) will be shortly called the Lyapunov solution. Exercise 13.11 Show that the Lyapunov solution 𝑋 satisfies 1 2𝜆𝑎 ≤ − , 𝛽
(13.12)
where 𝜆𝑎 is the spectral abscissa. Hint: compute 𝑦 ∗ 𝐵𝑦 on any eigenvector 𝑦 of 𝐴. If 𝐴 were normal, then (13.8) would simplify to ∥𝑒𝐴𝑡∥ = 𝑒𝜆𝑎 𝑡 as is immediately seen when 𝐴 is unitarily diagonalised, i.e. here the spectrum alone gives the full information on the decay. In this case for 𝐵 = 𝐼 (13.12) goes over into an equality because the Lyapunov solution reads 𝑋 = −(𝐴 + 𝐴∗ )−1 and 2𝜎(𝑋) = −1/ Re 𝜎(𝐴).
120
13 The Matrix Exponential
Each side of the estimate (13.12) is of a different nature: the right hand side depends on the scalar product in 𝛯 𝑛 (because it involves the adjoint of 𝐴) whereas the left hand side does not. Taking the matrix 𝐴1 = 𝐻 1/2 𝐴𝐻 −1/2 with 𝐻 positive definite (13.9) becomes 𝐴∗1 𝑋1 + 𝑋1 𝐴1 = −𝐵1 with
𝑋1 = 𝐻 −1/2 𝑋𝐻 −1/2 ,
𝐵1 = 𝐻 −1/2 𝐵𝐻 −1/2 .
This gives rise to a 𝛽1 which again satisfies (13.12). It is a non-trivial fact that by taking 𝐵 = 𝐼 and varying 𝐻 arbitrarily the value −1/𝛽1 comes arbitrarily close to 2𝜆𝑎 . The proof is rather simple, it is similar to the one that ∥𝐻 1/2 𝐴𝐻 −1/2 ∥ is arbitrarily close to spr(𝐴) if 𝐻 varies over the set of all positive definite matrices. Here it is. Since we are working with spectra and spectral norms we can obviously assume 𝐴 to be upper triangular. Set 𝐴′ = 𝐷𝐴𝐷−1 , 𝐷 = diag(1, 𝑀, . . . , 𝑀 𝑛−1 ), 𝑀 > 0. Then 𝐴′ = (𝑎′𝑖𝑗 ) is again upper triangular, has the same diagonal as 𝐴 and 𝑎′𝑖𝑗 = 𝑎𝑖𝑗 /𝑀 𝑗−𝑖 hence 𝐴′ → 𝐴0 = diag(𝑎11 , . . . , 𝑎𝑛𝑛 ) as 𝑀 → ∞. The matrix 𝐴0 is normal and the Lyapunov solution obviously depends continuously on 𝐴. This proves the statement. On the other hand we always have the lower bound ∥𝑒𝐴𝑡∥ ≥ 𝑒𝜆𝑎 𝑡 .
(13.13)
Indeed, let 𝜆 be the eigenvalue of 𝐴 for which Re 𝜆 = 𝜆𝑎 . Then 𝐴𝑥 = 𝜆𝑥 implies ∥𝑒𝐴𝑡 𝑥∥ = 𝑒𝜆𝑎 𝑡 ∥𝑥∥ and (13.13) follows. Exercise 13.12 Show that whenever the multiplicity 𝑝 of an eigenvalue is larger than the rank 𝑟 of the damping matrix 𝐶 then this eigenvalue must be defective. Express the highest non-vanishing power of the corresponding nilpotent by means of 𝑟 and 𝑝. Exercise 13.13 Replace the matrix 𝐶 in Example 12.22 by [ 𝐶=
𝑐0 00
]
Compute numerically the spectral abscissa for 2 < 𝑐 < 6 and find its minimum.
Chapter 14
The Quadratic Eigenvalue Problem
Thus far we have studied the damped system through its phase space matrix A by taking into account its J-Hermitian structure and the underlying indefinite metric. Now we come back to the fact that this matrix stems from a second order system which carries additional structure commonly called ‘the quadratic eigenvalue problem’. We study the spectrum of A and the behaviour of its exponential over time. A special class of so-called overdamped systems will be studied in some detail. If we set f = 0 in (3.2), and make the substitution y(t) = eλt y, y a constant vector, we obtain Ay = λy (14.1) and similarly, if in the homogeneous equation (1.1) we insert x(t) = eλt x, x constant we obtain (λ2 M + λC + K)x = 0 (14.2) which is called the quadratic eigenvalue problem, attached to (1.1), λ is an eigenvalue and x a corresponding eigenvector. We start here a detailed study of these eigenvalues and eigenvectors constantly keeping the connection with the corresponding phase-space matrix. Here, too, we can speak of the ‘eigenmode’ x(t) = eλt x but the physical appeal is by no means as cogent as in the undamped case (Exercise 2.1) because the proportionality is a complex one. Also, the general solution of the homogeneous equation (1.1) will not always be given as a superposition of the eigenmodes. Equations (14.1) and (14.2) are immediately seen to be equivalent via the substitution (for generality we keep on having complex M, C, K)
L∗1 x y= , K = L1 L∗1 , M = L2 L∗2 . λL∗2 x
K. Veseli´ c, Damped Oscillations of Linear Systems, Lecture Notes in Mathematics 2023, DOI 10.1007/978-3-642-21335-9 14, © Springer-Verlag Berlin Heidelberg 2011
(14.3)
121
122
14 The Quadratic Eigenvalue Problem
Equation (14.2) may be written as Q(λ)x = 0, where Q(·), defined as Q(λ) = λ2 M + λC + K
(14.4)
is the quadratic matrix pencil associated with (1.1). The solutions of the equation Q(λ)x = 0 are referred to as the eigenvalues and the eigenvectors of the pencil Q(·). For C = 0 the eigenvalue equation reduces to (2.3) with λ2 = −μ. Thus, to two different eigenvalues ±iω of (14.2) there corresponds only one linearly independent eigenvector. In this case the pencil can be considered as linear. Exercise 14.1 Show that in each of the cases 1. C = αM 2. C = βK the eigenvalues lie on a simple curve in the complex plane. Which are the curves? The set Xλ = {x : Q(λ)x = 0} is the eigenspace for the eigenvalue λ of the pencil Q(·), attached to (1.1). In fact, (14.3) establishes a one-to-one linear map from Xλ onto the eigenspace Yλ = {y ∈ C2n : Ay = λy}, in particular, dim(Yλ ) = dim(Xλ ). In other words, ‘the eigenproblems (14.1) and (14.2) have the same geometric multiplicity’. As we shall see soon, the situation is the same with the algebraic multiplicities. As is well known, the algebraic multiplicity of an eigenvalue of A is equal to its multiplicity as the root of the polynomial det(λI − A) or, equivalently, the dimension of the root space Eλ . With J from (3.7) we compute
I 0 JA − λJ = −1 −L−1 2 L1 /λ L2
−λ 0
2
0
λ M+λC+K λ
I −L∗1 L−∗ 2 /λ , (14.5) 0 L−∗ 2
where L1 , L2 are from (14.3). Thus, after some sign manipulations, −∗ 2 det(λI − A) = det(L−1 2 (λ M + λC + K)L2 ) = det Q(λ)/ det M.
Hence the roots of the equation det(λI − A) = 0 coincide with those of det(λ2 M + λC + K) = 0
14 The Quadratic Eigenvalue Problem
123
including their multiplicities. This is what is meant by saying that the algebraic multiplicity of an eigenvalue of A is equal to its algebraic multiplicity as an eigenvalue of Q(·). The situation is similar with the phase-space matrix A from (4.19). The linear relation x = Φy,
z1 = Ω 1 y 1 ,
z2 = Ω2 y2 ,
z3 = λy1
(14.6)
establishes a one-to-one correspondence between the eigenspace Zλ for A and the eigenspace Xλ for (14.2) (here, too, we allow complex Hermitian M, C, K). To compute a decomposition analogous to (14.5), we consider the inverse in (4.23): −1 1 Ω DΩ −1 + λ1 In F −1 (14.7) − JA + J = λ F∗ − λ1 Im Z 0 =W W ∗, 0 − λ1 Im where λ is real and
and
I −λF W = n 0 Im
λΩ −1 DΩ −1 + In + λ2 F F ∗ λ −1 Ω I 0 + λD + Ω 2 Ω −1 = λ2 n λ 0 0 Z=
Ω −1 ∗ Φ Q(λ)ΦΩ −1 . λ Now, using this, (14.7) as well as the equality =
det A = det(Ω)2 (det D22 )−1 (−1)n+m
(14.8)
we obtain J −1 − JA A = (det D22 )−1 (det Φ)2 det Q(λ), det(λI − A) = det −Jλ λ where Φ is from (4.7) and D22 from (4.8). From this it again follows that the algebraic multiplicity of an eigenvalue of A equals the algebraic multiplicity of the same eigenvalue of the pencil Q(·). That is why we have called the phase-space matrix a ‘linearisation’: the quadratic eigenvalue problem is replaced by one which is linear and equivalent to it.
124
14 The Quadratic Eigenvalue Problem
Let us now concentrate on the eigenvalue equation (14.2). It implies λ2 x∗ M x + λx∗ Cx + x∗ Kx = 0, x = 0
(14.9)
λ ∈ { p+ (x), p− (x) }
(14.10)
−x∗ Cx ± Δ(x) , p± (x) = 2x∗ M x
(14.11)
Δ(x) = (x∗ Cx)2 − 4x∗ Kxx∗ M x.
(14.12)
and hence with
The functions p± are homogeneous: p± (αx) = p± (x) for any scalar α and any vector x, both different from zero. If the discriminant Δ(x) is non-negative, then p± (x) is real and negative. If x is an eigenvector then by (14.10) the corresponding eigenvalue λ is also real and negative. Now for the corresponding y from (14.3) we have y∗ Jy = x∗ Kx − λ2 x∗ M x = 2x∗ Kx + λx∗ Cx
4x∗ M xx∗ Kx − (x∗ Cx)2 ± x∗ Cx Δ(x) = 2x∗ M x Δ(x) ± x∗ Cx Δ(x) = ∓λ Δ(x), = ∗ 2x M x
(14.13)
where we have used the identity (14.9) and the expressions (14.11) and (14.12). Since −λ is positive, we see that the eigenvector y of A is J-positive or J-negative according to the sign taken in (14.11). The same is true of A from (4.19) and z from (14.6). Indeed, with J from (4.22), we have z ∗ Jz = y1T Ω1 y1 + y2T Ω2 y2 − λ2 y1T y1 = y T Ωy − λ2 y T
Im 0 y = xT Kx − λ2 xT M x 0 0
as in (14.13). We summarise: Proposition 14.2 The following statements are equivalent 1. λ is an eigenvalue of (14.2) and every corresponding eigenvector satisfies Δ(x) > 0. 2. λ is a J-definite eigenvalue of the phase-space matrix A from either (3.3) or (4.19).
14 The Quadratic Eigenvalue Problem
125
In this case λ is J-positive (J-negative) if and only if λ = p+(x) (λ = p− (x)) for any corresponding eigenvector x. The appellation positive/negative/definite type will be naturally extended to the eigenvalues of Q(·). The reader will notice that we do not use the term ‘overdamped’ for such eigenvalues (as we did for the one-dimensional oscillator in Example 9.3). The overdampedness property in several dimensions is stronger than that as will be seen in the sequel. If all eigenvalues of (14.2) are real and simple then they are of definite type (cf. Exercise 12.7). We call the family Q(·) (or, equivalently, the system (1.1)) overdamped, if there is a μ such that Q(μ) = μ2 M + μC + K is negative definite. Proposition 14.3 For λ negative we have ι+ (JA − λJ) = ι− (Q(λ)) + n and consequently ι− (JA − λJ) = ι+ (Q(λ)). for A, J from (3.3), (3.7), respectively. In particular, Q(λ) is negative definite, if and only if JA − λJ is positive definite. Proof. The assertion immediately follows from (14.5) and the Sylvester inertia theorem (Theorem 5.7). Q.E.D. Corollary 14.4 Let Q(·) be overdamped with eigenvalues − + + λ− n− ≤ · · · ≤ λ1 < λ1 ≤ · · · ≤ λn+ + + and the definiteness interval (λ− 1 , λ1 ). Then for any λ > λ1 , with Q(λ) non-singular the quantity ι+ (Q(λ)) is equal to the number of the J-positive eigenvalues less than λ (and similarly for the J-negative eigenvalues).
Proof. The statement follows from Proposition 14.3 and Theorem 10.7. Q.E.D. Thus, the overdampedness of the system M, C, K is equivalent to the definitisability of its phase-space matrix. By Theorem 10.6 the eigenvalues are split into two groups: the J-positive eigenvalues on the right and the J-negative ones on the left. They are divided by the definiteness interval which will go under the same name for the pencil Q(·) as well. The overdampedness condition means x∗ M xλ2 + x∗ Cxλ + x∗ Kx ≡ x∗ M x(λ − p− (x))(λ − p+ (x)) < 0
(14.14)
126
14 The Quadratic Eigenvalue Problem
for any fixed λ from the definiteness interval and for all non-vanishing x, that is p− (x) < λ, p+ (x) > λ, in particular, p− (x) < p+ (x), or equivalently Δ(x) > 0
(14.15)
for every x = 0. Prima facie the condition (14.15) is weaker than (14.14). It is a non-trivial fact that the two are indeed equivalent as will be seen presently. The overdampedness is not destroyed, if we increase the damping and/or decrease the mass and the stiffness, that is, if we replace M, C, K by M , C , K with x∗ M x ≤ x∗ M x, x∗ C x ≥ x∗ Cx, x∗ K x ≤ x∗ Kx for all x. In this situation we say that the system M , C , K is more viscous than M, C, K. The viscosity is, in fact, an order relation on the set of all damped systems. We now formulate a theorem collecting together several equivalent definitions of the overdampedness. Theorem 14.5 The following properties are equivalent: (i) The system (1.1) is overdamped. (ii) The discriminant Δ(x) from (14.12) is positive for all non-vanishing x ∈ Cn . (iii) The pair JA, J or, equivalently, S, T from (11.6) is definitisable. Proof. The relations (i) ⇐⇒ (iii), (i) =⇒ (ii) are obvious. To prove (ii) =⇒ (i) we will use the following, also obvious, facts: • The property (ii) is preserved under increased viscosity. • The property (i) is preserved under increased viscosity, moreover the new definiteness interval contains the old one. • The set of all semidefinite matrices is closed. • The set of all overdamped systems is open, that is, small changes in the matrices M, C, K do not destroy overdampedness. Suppose, on the contrary, that (ii) =⇒ (i) does not hold. Then there would exist an 0 ≥ 1 such that the pencils Q(λ, ) = λ2 M + λ C + K are overdamped for all > 0 but not for = 0 (obviously the overdampedness holds, if is large enough). Since the corresponding phase-space matrix A ) depends continuously – and the definiteness intervals monotonically – on
14 The Quadratic Eigenvalue Problem
127
> 0 there is a real point λ = λ0 , contained in all these intervals and this is an eigenvalue of A0 – otherwise JA0 − λ0 J would be not only positive semidefinite (which it must be), but positive definite which is precluded. Thus JA0 − λ0 J is singular positive semidefinite, but by (ii) the eigenvalue λ0 is of definite type. By Theorem 10.11 the pair JA0 , J is definitisable i.e. JA0 − λJ is positive definite for some λ in the neighbourhood of λ0 . This proves (ii) =⇒ (i). Q.E.D. Exercise 14.6 Let the system M, C, K be overdamped and let C be perturbed into C = C + δC with |xT δCx| ≤ xT Cx
for all x,
1 xT Cx .
< 1 − , d = min √ x 2 xT M xxT Kx d Show that the system M, C , K is overdamped. Exercise 14.7 Let M, C, K be real symmetric positive definite matrices. Show that Δ(x) > 0 for all real x = 0 is equivalent to Δ(x) > 0 for all complex x = 0. Exercise 14.8 Let the damping be proportional: C = αM + βK, α, β ≥ 0. Try to tell for which choices of α, β the system is overdamped. Show that this is certainly the case, if αβ > 1. Exercise 14.9 Prove the following statement. If μ is negative and Q(μ) nonsingular then A has at least 2(n − ι− (Q(μ))) real eigenvalues (counting multiplicities). Exercise 14.10 Prove the following statement. Let Q(·) have p eigenvalues λ1 , . . . , λp (counting multiplicities). If all these eigenvalues are of positive type then corresponding eigenvectors x1 , . . . , xp can be chosen as linearly independent. Hint: Use the fact that the J-Gram matrix X ∗ JX for the corresponding phase space eigenvectors
L∗1 xj yj = λj L∗2 xj
is positive definite. Exercise 14.11 Prove the identity (14.8). Check this identity on the system from Example 4.5.
Chapter 15
Simple Eigenvalue Inclusions
Although, as we have seen, the eigenvalues alone may not give sufficient information on the behaviour of the damped system, knowledge of them may still be useful. In this chapter we will derive general results on the location of the eigenvalues of a damped system. In fact, the parameters obtained in such ‘bounds’ or ‘inclusions’ may give direct estimates for the matrix exponential which are often more useful than (13.5) and (13.6). An example is the numerical range or the field of values 𝑟(𝐴) = {𝑥∗ 𝐴𝑥 : 𝑥 ∈ ℂ𝑛 , ∥𝑥∥ = 1} which contains 𝜎(𝐴). This is seen, if the eigenvalue equation 𝐴𝑥 − 𝜆𝑥 = 0 is premultiplied with 𝑥∗ which gives 𝜆=
𝑥∗ 𝐴𝑥 . 𝑥∗ 𝑥
Now (3.9) can be strengthened to 𝑑 𝑑 ∥𝑦∥2 = 𝑦 ∗ 𝑦 = 𝑦˙ ∗ 𝑦 + 𝑦 ∗ 𝑦˙ = 𝑦 ∗ (𝐴∗ + 𝐴)𝑦 𝑑𝑡 𝑑𝑡 = 2 Re 𝑦 ∗ 𝐴𝑦 ≤ 2 max Re(𝑟(𝐴))∥𝑦∥2 = 𝜆𝑎 ∥𝑦∥2 . By integrating the inequality 𝑑 ∥𝑦∥2 𝑑𝑡 ∥𝑦∥2
we obtain
=
𝑑 1 ≤ 2 max Re(𝑟(𝐴)) 𝑑𝑡 ∥𝑦∥2
∥𝑦∥2 ≤ 𝑒2 max Re(𝑟(𝐴)) ∥𝑦(0)∥2 .
K. Veseli´ c, Damped Oscillations of Linear Systems, Lecture Notes in Mathematics 2023, DOI 10.1007/978-3-642-21335-9 15, © Springer-Verlag Berlin Heidelberg 2011
129
130
15 Simple Eigenvalue Inclusions
Similarly, the inequality 𝑑 ∥𝑦∥2 = 2 Re 𝑦 ∗ 𝐴𝑦 ≥ 2 min Re(𝑟(𝐴))∥𝑦∥2 𝑑𝑡 implies
∥𝑦∥2 ≥ 𝑒2 min Re(𝑟(𝐴)) ∥𝑦(0)∥2 .
Altogether we obtain 𝑒min Re(𝑟(𝐴))𝑡 ≤ ∥𝑒𝐴𝑡∥ ≤ 𝑒max Re(𝑟(𝐴))𝑡 .
(15.1)
This is valid for any matrix 𝐴. If 𝐴 is our phase-space matrix then max Re(𝑟(𝐴)) = 0 which leads to the known estimate ∥𝑒𝐴𝑡 ∥ ≤ 1 whereas the left hand side inequality in (15.1) yields ∥𝑒𝐴𝑡 ∥ ≥ 𝑒−∥𝐶∥𝑡 . The decay information obtained in this way is rather poor. A deeper insight comes from the quadratic eigenvalue equation (14.2). We thus define the numerical range of the system 𝑀, 𝐶, 𝐾, also called the quadratic numerical range as 𝑤(𝑀, 𝐶, 𝐾) = {𝜆 ∈ ℂ : 𝜆2 𝑥∗ 𝑀 𝑥 + 𝜆𝑥∗ 𝐶𝑥 + 𝑥∗ 𝐾𝑥 = 0, 𝑥 ∈ ℂ𝑛 ∖ {0}} or, equivalently,
𝑤(𝑀, 𝐶, 𝐾) = ℛ(𝑝+ ) ∪ ℛ(𝑝− ),
where the symbol ℛ denotes the range of a map and 𝑝± is given by (14.11); for convenience we set a non-real 𝑝+ to have the positive imaginary part. By construction, the quadratic numerical range contains all eigenvalues.1 Due to the obvious continuity and homogeneity of the functions 𝑝± the set 𝑤(𝑀, 𝐶, 𝐾) is compact. (That is, under our standard assumption that 𝑀 be positive definite. If 𝑀 is singular, then 𝑤(𝑀, 𝐶, 𝐾) is unbounded. This has to do with the fact that the ‘missing’ eigenvalues in this case can be considered as infinite.) The non-real part of 𝑤(𝑀, 𝐶, 𝐾) is confined to certain regions of the left half plane: Proposition 15.1 Let 𝜆 ∈ 𝑤(𝑀, 𝐶, 𝐾) be non real. Then
1 This
𝑐− ≤ −𝑅𝑒𝜆 ≤ 𝑐+
(15.2)
𝜔1 ≤ ∣𝜆∣ ≤ 𝜔𝑛 ,
(15.3)
is a special case of the so called polynomial numerical range, analogously defined for general matrix polynomials.
15 Simple Eigenvalue Inclusions
where 𝑐− = min
131
𝑥𝑇 𝐶𝑥 𝑥𝑇 𝐶𝑥 , 𝑐 = sup , + 2𝑥𝑇 𝑀 𝑥 2𝑥𝑇 𝑀 𝑥
(15.4)
𝑥𝑇 𝐾𝑥 𝑥𝑇 𝐾𝑥 2 , 𝜔 = sup . (15.5) 𝑛 𝑥𝑇 𝑀 𝑥 𝑥𝑇 𝑀 𝑥 Here infinity as sup value is allowed, if 𝑀 has a non-trivial null space and only those 𝑥 are taken for which 𝑥𝑇 𝑀 𝑥 > 0. In particular, any non-real eigenvalue 𝜆 satisfies (15.2) and (15.3). 𝜔12 = min
Proof. If 𝜆 is non-real, then 𝜆=
−𝑥∗ 𝐶𝑥 ± 𝑖
√ 4𝑥∗ 𝐾𝑥𝑥∗ 𝑀 𝑥 − (𝑥∗ 𝐶𝑥)2 2𝑥∗ 𝑀 𝑥
and Re 𝜆 = −
𝑥∗ 𝐶𝑥 2𝑥∗ 𝑀 𝑥
from which (15.2) follows. Similarly, ∣𝜆∣2 =
𝑥∗ 𝐾𝑥 𝑥∗ 𝑀 𝑥
from which (15.3) follows. Q.E.D. Note that for 𝑀 non-singular the suprema in (15.4), (15.5) turn to maxima. The eigenvalue inclusion obtained in Proposition 15.1 can be strengthened as follows. Theorem 15.2 Any eigenvalue 𝜆 outside of the region (15.2), (15.3) is real and of definite type. More precisely, any 𝜆 < max{−𝑐− , −𝜔1 } is of negative type whereas any 𝜆 > min{−𝑐+ , −𝜔𝑛 } is of positive type. Proof. The complement of the region (15.2), (15.3) contains only real negative eigenvalues 𝜆. Hence for any eigenpair 𝜆, 𝑥 of (14.2) we have 𝛥(𝑥) ≥ 0 and 𝜆 ∈ {𝑝+ (𝑥), 𝑝− (𝑥)}. Suppose 𝜆 = 𝑝+ (𝑥) < −𝑐− that is, 𝑥∗ 𝐶𝑥 − ∗ + 2𝑥 𝑀 𝑥 Then
√
(
𝑥∗ 𝐶𝑥 2𝑥∗ 𝑀 𝑥
)2 −
𝑥∗ 𝐾𝑥 𝑦 ∗ 𝐶𝑦 < − max . 𝑦 2𝑦 ∗ 𝑀 𝑦 𝑥∗ 𝑀 𝑥
𝑥∗ 𝐶𝑥 𝑦 ∗ 𝐶𝑦 > max ∗ , ∗ 𝑦 2𝑦 𝑀 𝑦 2𝑥 𝑀 𝑥
132
15 Simple Eigenvalue Inclusions
a contradiction. Thus, 𝜆 ∕= 𝑝+ (𝑥) and 𝜆 = 𝑝− (𝑥), hence 𝛥(𝑥) > 0 and 𝜆 is of negative type. The situation with 𝜆 = 𝑝− (𝑥) > −𝑐+ is analogous. Suppose now 𝜆 = 𝑝+ (𝑥) < −𝜔𝑛 that is ( √ )( ) √ √ 𝑥∗ 𝐶𝑥 𝑥∗ 𝐾𝑥 𝑥∗ 𝐶𝑥 𝑥∗ 𝐾𝑥 𝑥∗ 𝐶𝑥 𝑦 ∗ 𝐾𝑦 ⎷ + − < ∗ − max ∗ ∗ ∗ ∗ ∗ 𝑦 𝑦 𝑀𝑦 2𝑥 𝑀 𝑥 𝑥 𝑀𝑥 2𝑥 𝑀 𝑥 𝑥 𝑀𝑥 2𝑥 𝑀 𝑥 𝑥∗ 𝐶𝑥 ≤ ∗ − 2𝑥 𝑀 𝑥
√
𝑥∗ 𝐾𝑥 . 𝑥∗ 𝑀 𝑥
So, we would have 𝑥∗ 𝐶𝑥 + 2𝑥∗ 𝑀 𝑥
√
𝑥∗ 𝐾𝑥 𝑥∗ 𝐶𝑥 < ∗ − ∗ 𝑥 𝑀𝑥 2𝑥 𝑀 𝑥
√
𝑥∗ 𝐾𝑥 , 𝑥∗ 𝑀 𝑥
a contradiction. Hence 𝜆 ∕= 𝑝+ (𝑥), 𝜆 = 𝑝− (𝑥), 𝛥(𝑥) > 0 and by Proposition 14.2 𝜆 is of negative type. Suppose −𝜔1 < 𝜆 = 𝑝− (𝑥). Then 𝑥∗ 𝐶𝑥 1 =− ∗ + 𝑝− (𝑥) 2𝑥 𝐾𝑥
√(
𝑥∗ 𝐶𝑥 2𝑥∗ 𝐾𝑥
)2 −
𝑦∗ 𝑀 𝑦 𝑥∗ 𝑀 𝑥 < − max . 𝑦 𝑥∗ 𝐾𝑥 𝑦 ∗ 𝐾𝑦
or again ( √ )( ) √ √ 𝑥∗ 𝐶𝑥 𝑥∗ 𝑀 𝑥 𝑥∗ 𝐶𝑥 𝑥∗ 𝑀 𝑥 𝑥∗ 𝐶𝑥 𝑦∗𝑀 𝑦 ⎷ + − < ∗ − max ∗ . ∗ ∗ ∗ ∗ 𝑦 2𝑥 𝐾𝑥 𝑥 𝐾𝑥 2𝑥 𝐾𝑥 𝑥 𝐾𝑥 2𝑥 𝐾𝑥 𝑦 𝐾𝑦 This leads to a contradiction similarly as before. Q.E.D. Remark 15.3 Another set of eigenvalue inclusions is obtained if we note that the adjoint eigenvalue problem (𝜇2 𝐾 + 𝜇𝐶 + 𝑀 )𝑥 = 0 has the same eigenvectors as (14.2) and the inverse eigenvalues 𝜇 = 1/𝜆. Analogously to (15.2) non-real eigenvalues 𝜆 are contained in the set difference of two circles
15 Simple Eigenvalue Inclusions
133
{
( )2 ( )} 1 1 2 2 𝜇1 + 𝑖𝜇2 : 𝜇1 + + 𝜇2 ≤ ∖ 2𝛾− 2𝛾−
{
( )2 ( )2 } 1 1 2 𝜇1 + 𝑖𝜇2 : 𝜇1 + + 𝜇2 ≤ 2𝛾+ 2𝛾+
with
𝑥∗ 𝐶𝑥 𝑥∗ 𝐶𝑥 , 𝛾− = min ∗ ∗ 𝑥 2𝑥 𝐾𝑥 𝑥 2𝑥 𝐾𝑥 (the analog of (15.3) gives nothing new). 𝛾+ = max
Proposition 15.4 Let 𝑀 be non-singular and 𝑛 eigenvalues (counting multiplicities) of 𝐴 are placed on one side of the interval [−𝑐+ , −𝑐− ] ∩ [−𝜔𝑛 , −𝜔1 ]
(15.6)
with 𝑐± as in Proposition 15.1. Then the system is overdamped. + Proof. Let 𝑛 eigenvalues 𝜆+ 1 , . . . , 𝜆𝑛 be placed, say, in
(− min{𝑐− , 𝜔1 }, 0). Then their eigenvectors are 𝐽-positive and all these eigenvalues are 𝐽-positive. Therefore to them there correspond 𝑛 eigenvectors of 𝐴: 𝑦1+ , . . . , 𝑦𝑛+ with 𝑦𝑗∗ 𝐽𝑦𝑘 = 𝛿𝑗𝑘 and they span a 𝐽-positive subspace 𝒳+ , invariant for 𝐴. Then the 𝐽-orthogonal complement 𝒳− is also invariant for 𝐴 and is 𝐽-negative (note that 𝑛 = 𝜄+ (𝐽) = 𝜄− (𝐽)) hence other eigenvalues are 𝐽-negative and they cannot belong to the interval (15.6). Thus we have the situation from Theorem 10.6. So there is a 𝜇 such that 𝐽𝐴 − 𝜇𝐽 is positive definite and the system is overdamped. Q.E.D. The systems having one or more purely imaginary eigenvalues in spite of the damping have special properties: Proposition 15.5 An eigenvalue 𝜆 in (14.2) is purely imaginary, if and only if for the corresponding eigenvector 𝑥 we have 𝐶𝑥 = 0. Proof. If 𝐶𝑥 = 0 then 𝜆 is obviously purely imaginary. Conversely, let 𝜆 = 𝑖𝜔 with 𝜔 real i.e. −𝜔 2 𝑀 𝑥 + 𝑖𝜔𝐶𝑥 + 𝐾𝑥 = 0 hence
−𝜔 2 𝑥∗ 𝑀 𝑥 + 𝑖𝜔𝑥∗ 𝐶𝑥 + 𝑥∗ 𝐾𝑥 = 0.
By taking the imaginary part we have 𝑥∗ 𝐶𝑥 = 0 and, by the positive semidefiniteness of 𝐶, 𝐶𝑥 = 0. Q.E.D.
134
15 Simple Eigenvalue Inclusions
The presence of purely imaginary eigenvalues allows to split the system into an orthogonal sum of an undamped system and an exponentially decaying one. We have Proposition 15.6 Whenever there are purely imaginary eigenvalues there is a non-singular 𝛷 such that ] 0 𝐾1 (𝜆) 𝛷 𝑄(𝜆)𝛷 = , 0 𝐾2 (𝜆) ∗
[
𝐾1 (𝜆) = 𝜆2 𝑀1 + 𝐾1 ,
𝐾2 (𝜆) = 𝜆2 𝑀2 + 𝜆𝐶2 + 𝐾2 ,
where the system 𝑀2 , 𝐶2 , 𝐾2 has no purely imaginary eigenvalues. Proof. Take any eigenvector of 𝑄(⋅) for an eigenvalue 𝑖𝜔. Then by completing it to an 𝑀 -orthonormal basis in 𝛯 𝑛 we obtain a matrix 𝛷1 such that 𝛷∗1 𝑄(𝜆)𝛷1 =
[
0 𝜆2 − 𝜔 2 ˆ 0 𝐾(𝜆)
]
and, if needed, repeat the procedure on ˆ ˆ + 𝜆𝐶ˆ + 𝐾. ˆ 𝐾(𝜆) = 𝜆2 𝑀 and finally set 𝛷 = 𝛷1 𝛷2 ⋅ ⋅ ⋅. Q.E.D. Corollary 15.7 The phase-space matrix 𝐴 from (3.3) or (4.19) is asymptotically stable, if and only if the form 𝑥∗ 𝐶𝑥 is positive definite on any eigenspace of the undamped system. Any system which is not asymptotically stable is congruent to an orthogonal sum of an undamped system and an asymptotically stable one. Proof. According to Proposition 15.5 𝐴 is asymptotically stable, if and only if 𝐶𝑥 ∕= 0 for any non-vanishing 𝑥 from any undamped eigenspace. By the positive semidefiniteness of 𝐶 the relation 𝐶𝑥 ∕= 0 is equivalent to 𝑥∗ 𝐶𝑥 > 0. The last statement follows from Proposition 15.6. Q.E.D. If 𝑖𝜔 is an undamped eigenvalue of multiplicity 𝑚 then 𝐶 must have rank at least 𝑚 in order to move it completely into the left half-plane. (That eigenvalue may then, of course, split into several complex eigenvalues.) In particular, if 𝐶 has rank one then the asymptotic stability can be achieved, only if all undamped eigenvalues are simple. In spite of their elegance and intuitiveness the considerations above give no further information on the exponential decay of 𝑒𝐴𝑡 . Exercise 15.8 Show that the system from Example 1.1 is asymptotically stable, if any of 𝑐𝑖 or 𝑑𝑖 is different from zero.
Chapter 16
Spectral Shift
Spectral shift is a simple and yet powerful tool to obtain informations on the decay in time of a damped system. If in the differential equation 𝑦˙ = 𝐴𝑦 we substitute
𝑦 = 𝑒𝜇𝑡 𝑢,
𝑦˙ = 𝜇𝑒𝜇𝑡 𝑢 + 𝑒𝜇𝑡 𝑢˙
then we obtain the ‘spectral-shifted’ system ˜ 𝑢˙ = 𝐴𝑢,
𝐴˜ = 𝐴 − 𝜇𝐼.
Now, if we know that 𝐴˜ has an exponential which is bounded in time and 𝜇 < 0 this property would yield an easy decay bound ˜
∥𝑒𝐴𝑡∥ ≤ 𝑒𝜇𝑡 sup ∥𝑒𝐴𝑡 ∥. 𝑡≥0
In the case of our phase-space matrix 𝐴 it does not seem to be easy to find ˜ or to estimate the quantity sup𝑠≥0 ∥𝑒𝐴𝑡 ∥ just from the spectral properties of 𝐴. However, if an analogous substitution is made on the second order system (1.1) the result is much more favourable. So, by setting 𝑥 = 𝑒𝜇𝑡 𝑧 and using 𝑥˙ = 𝜇𝑒𝜇𝑡 𝑧 + 𝑒𝜇𝑡 𝑧˙ 𝑥 ¨ = 𝜇2 𝑒𝜇𝑡 𝑧 + 2𝜇𝑒𝜇𝑡 𝑧˙ + 𝑒𝜇𝑡 𝑧¨ the system (1.1) (with 𝑓 ≡ 0) goes over into 𝑀 𝑧¨ + 𝐶(𝜇)𝑧˙ + 𝑄(𝜇)𝑧 = 0 K. Veseli´ c, Damped Oscillations of Linear Systems, Lecture Notes in Mathematics 2023, DOI 10.1007/978-3-642-21335-9 16, © Springer-Verlag Berlin Heidelberg 2011
135
136
16 Spectral Shift
with 𝑄(𝜇) from (14.4) and 𝐶(𝜇) = 𝐾 ′ (𝜇) = 2𝜇𝑀 + 𝐶. As long as 𝐶(𝜇), 𝑄(𝜇) stay positive definite this is equivalent to the phasespace representation ˆ 𝑤˙ = 𝐴𝑤,
𝐴ˆ =
[
0
−𝑀 −1/2 𝑄(𝜇)1/2 [ 𝑤=
] 𝑄(𝜇)1/2 𝑀 −1/2 , −𝑀 −1/2 𝐶(𝜇)𝑀 −1/2
(16.1)
] 𝑄(𝜇)1/2 𝑧 , 𝑀 1/2 𝑧˙
ˆ
whose solution is given by 𝑒𝐴𝑡 . We now connect the representations (16.1) and (3.2). We have 𝑦1 = 𝐾 1/2 𝑥 = 𝑒𝜇𝑡 𝐾 1/2 𝑧 = 𝜇𝑒𝜇𝑡 𝐾 1/2 𝑄(𝜇)−1/2 𝑤1 , 𝑦2 = 𝑀 1/2 𝑥˙ = 𝜇𝑒𝜇𝑡 𝑀 1/2 𝑄(𝜇)−1/2 𝑤1 + 𝑒𝜇𝑡 𝑤2 . Thus,
𝑦 = 𝑒𝜇𝑡 ℒ(𝜇)𝑤, ] [ 𝐾 1/2 𝑄(𝜇)−1/2 0 . ℒ(𝜇) = 𝜇𝑀 1/2 𝑄(𝜇)−1/2 𝐼
This, together with the evolution equations for 𝑦, 𝑤 gives ℒ(𝜇)𝐴ˆ = (𝐴 − 𝜇𝐼) ℒ(𝜇) or, equivalently
(16.2)
ˆ
𝑒𝐴𝑡 = ℒ−1 (𝜇)𝑒𝐴+𝜇𝑡𝐼 ℒ(𝜇).
Hence an elegant decay estimate ∥𝑒𝐴𝑡 ∥ ≤ ∥ℒ(𝜇)∥∥ℒ(𝜇)−1 ∥𝑒𝜇𝑡 .
(16.3)
This brings exponential decay only, if there are such 𝜇 < 0 for which 𝐶(𝜇), 𝑄(𝜇) stay positive definite, in fact, 𝐶(𝜇) may be only positive semidefinite. By continuity, such 𝜇 will exist, if and only if 𝐶 is positive definite. That is, positive definite 𝐶 insures the exponential decay. The next task is to establish the optimal decay factor 𝜇 in (16.3). We set 𝛾 = max Re 𝑝+ (𝑥) = max Re 𝑤(𝑀, 𝐶, 𝐾), 𝑥∕=0
this maximum is obtained because 𝑤(𝑀, 𝐶, 𝐾) is compact.
16 Spectral Shift
137
Theorem 16.1 Let all of 𝑀, 𝐶, 𝐾 be all positive definite. Then 𝛾 ≥ − inf 𝑥
𝑥∗ 𝐶𝑥 , 2𝑥∗ 𝑀 𝑥
(16.4)
moreover, 𝛾 is the infimum of all 𝜇 for which both 𝑄(𝜇) and 𝐶(𝜇) are positive definite. Proof. The relation (16.4) is obvious. To prove the second assertion take any 𝜇 > 𝛾. Then 𝑥∗ 𝑄(𝜇)𝑥 (16.5) = 𝑥∗ 𝑀 𝑥(𝜇 − 𝑝− (𝑥))(𝜇 − 𝑝+ (𝑥)) ≥ 𝑥∗ 𝑀 𝑥(𝜇 − Re 𝑝+ (𝑥))2 ≥ 𝑥∗ 𝑀 𝑥(𝜇 − 𝛾)2 (note that 𝑝− (𝑥) ≤ 𝑝+ (𝑥) whenever 𝛥(𝑥) ≥ 0). Thus, 𝑄(𝜇) is positive definite. By (16.4) 𝐶(𝜇) is positive definite also. By continuity, both 𝑄(𝛾) and 𝐶(𝛾) are positive semidefinite. Conversely, suppose that, say, 𝐶(𝛾) is positive definite; we will show that 𝑄(𝛾) must be singular. Indeed, take a unit vector 𝑥0 such that Re 𝑝+ (𝑥0 ) = 𝛾. Then by inserting in (16.5) 𝜇 = 𝛾 we obtain 𝑥∗ 𝑄(𝛾)𝑥0 = 0, i.e. 𝑄(𝛾) is singular. Q.E.D. The previous theorem gives a simple possibility to compute 𝛾 by iterating the Cholesky decomposition of 𝑄(𝜇) on the interval − inf 𝑥
𝑥∗ 𝐶𝑥 ≤𝜇 0, 𝑗
(19.8)
min 2𝜔𝑗 − ∥𝐷∥ > 0, 𝑗
with 𝑅𝑗 from (19.4) and min 𝑗
2 − ∥𝛺 −1 𝐷𝛺 −1 ∥ > 0 𝜔𝑗
precludes real eigenvalues of (14.2). Solution. If an eigenvalue 𝜆 is real then according to Exercise 19.3 we must have 𝜆2 + 𝜔𝑗2 ≤ ∣𝜆∣𝑅𝑗 for some 𝑗 and this implies 𝑅𝑗2 − 4𝜔𝑗2 ≥ 0. Thus, under (19.8) no real eigenvalues can occur (and similarly in other two cases). The just considered undamped approximation was just a prelude to the main topic of this chapter, namely the modal approximation. The modally damped systems are so much simpler than the general ones that practitioners often substitute the true damping matrix by some kind of ‘modal approximation’. Most typical such approximations in use are of the form 𝐶𝑝𝑟𝑜𝑝 = 𝛼𝑀 + 𝛽𝐾,
(19.9)
where 𝛼, 𝛽 are chosen in such a way that 𝐶𝑝𝑟𝑜𝑝 be in some sense as close as possible to 𝐶, for instance, Tr [(𝐶 − 𝛼𝑀 − 𝛽𝐾)𝑊 (𝐶 − 𝛼𝑀 − 𝛽𝐾)] = min, where 𝑊 is some convenient positive definite weight matrix. This is a proportional approximation. In general such approximations may go quite astray and yield thoroughly false predictions. We will now assess them in a more systematic way. A modal approximation to the system (1.1) is obtained by first representing it in modal coordinates by the matrices 𝐷, 𝛺 and then by replacing 𝐷 by its diagonal part 𝐷 0 = diag(𝑑11 , . . . , 𝑑𝑛𝑛 ). (19.10) The off-diagonal part
𝐷′ = 𝐷 − 𝐷0
(19.11)
19 Modal Approximation
149
is considered a perturbation. Again we can work in the phase-space or with the original quadratic eigenvalue formulation. In the first case we can make perfect shuffling to obtain [ 𝐴 = (𝐴𝑖𝑗 ),
𝐴𝑖𝑖 =
] 0 𝜔𝑖 , −𝜔𝑖 −𝑑𝑖𝑖
[ 𝐴𝑖𝑗 =
0 0 0 −𝑑𝑖𝑗
] (19.12)
𝐴0 = diag(𝐴11 , . . . , 𝐴𝑛𝑛 ). So, for 𝑛 = 3
Then
⎡
0 ⎢ −𝜔 1 ⎢ ⎢ ⎢ 0 𝐴=⎢ ⎢ 0 ⎢ ⎣ 0 0
0 0 0 𝜔1 −𝑑11 0 −𝑑12 0 0 0 𝜔2 0 −𝑑12 −𝜔2 −𝑑22 0 0 0 0 0 −𝑑13 0 −𝑑23 −𝜔3
⎤ 0 −𝑑13 ⎥ ⎥ ⎥ 0 ⎥ ⎥. −𝑑23 ⎥ ⎥ 𝜔3 ⎦ −𝑑33
∥(𝐴0 − 𝜆𝐼)−1 ∥−1 = max ∥(𝐴𝑗𝑗 − 𝜆𝐼𝑗 )−1 ∥−1 . 𝑗
Even for 2 × 2-blocks any common norm of (𝐴𝑗𝑗 − 𝜆𝐼𝑗 )−1 seems complicated to express in terms of disks or other simple regions, unless we diagonalise each 𝐴𝑗𝑗 as [
] 𝜆𝑗+ 0 −1 𝑆𝑗 𝐴𝑗𝑗 𝑆𝑗 = , 0 𝜆𝑗−
𝜆𝑗± =
−𝑑𝑗𝑗 ±
√ 𝑑2𝑗𝑗 − 4𝜔𝑗2 2
.
(19.13)
As we know from Example 9.3 we have √ 𝜅(𝑆𝑗 ) =
1 + 𝜇2𝑗 , ∣1 − 𝜇2𝑗 ∣
𝜇𝑗 =
𝑑𝑗𝑗 . 2𝜔𝑗
Set 𝑆 = diag(𝑆11 , . . . , 𝑆𝑛𝑛 ) and 𝐴′ = 𝑆 −1 𝐴𝑆 = 𝐴′0 + 𝐴′′ , then
𝐴′′ = 𝑆 −1 (𝐴 − 𝐴0 )𝑆
𝐴′0 = diag(𝜆1± , . . . , 𝜆𝑛± ), 𝐴′′𝑗𝑘 = 𝑆𝑗−1 𝐴𝑗𝑘 𝑆𝑘 .
Now the general perturbation bound (19.2), applied to 𝐴′0 , 𝐴′′ , gives 𝜎(𝐴) ⊆ ∪𝑗,± {𝜆 : ∣𝜆 − 𝜆𝑗± ∣ ≤ 𝜅(𝑆)∥𝐷′ ∥}.
(19.14)
150
19 Modal Approximation
There is a related ‘Gershgorin-type bound’ 𝜎(𝐴) ⊆ ∪𝑗,± {𝜆 : ∣𝜆 − 𝜆𝑗± ∣ ≤ 𝜅(𝑆𝑗 )𝑟𝑗 } with 𝑟𝑗 =
𝑛 ∑
(19.15)
∥𝑑𝑗𝑘 ∥.
(19.16)
𝑘=1 𝑘∕=𝑗
To show this we replace the spectral norm ∥ ⋅ ∥ in (19.1) by the norm ∥∣ ⋅ ∥∣1 , defined as ∑ ∥∣𝐴∥∣1 := max ∥𝐴𝑘𝑗 ∥, 𝑗
𝑘
where the norms on the right hand side are spectral. Thus, (19.1) will hold, if ∑ max ∥(𝐴 − 𝐴0 )𝑘𝑗 ∥∥(𝐴𝑗𝑗 − 𝜆𝐼𝑗 )−1 ∥ < 1. 𝑗
𝑘
Taking into account the equality { ∥(𝐴 − 𝐴0 )𝑘𝑗 ∥ =
∣𝑑𝑘𝑗 ∣, 𝑘 = ∕ 𝑗 0𝑘=𝑗
𝜆 ∈ 𝜎(𝐴) implies 𝑟𝑗 ≥ ∥(𝐴𝑗𝑗 − 𝜆𝐼)−1 ∥ ≥
min{∣𝜆 − 𝜆𝑗+ ∣, ∣𝜆 − 𝜆𝑗− ∣} 𝜅(𝑆𝑗 )
and this is (19.15). The bounds (19.14) and (19.15) are poor whenever the modal eigenvalue approximation is close to a critically damped eigenvalue. Better bounds are expected, if we work directly with the quadratic eigenvalue equation. The inverse (𝜆2 𝐼 + 𝜆𝐷 + 𝛺 2 )−1 = (𝜆2 𝐼 + 𝜆𝐷 0 + 𝛺 2 )−1 (𝐼 + 𝜆𝐷 ′ (𝜆2 𝐼 + 𝜆𝐷 0 + 𝛺 2 )−1 )−1 exists, if
∥𝐷′ (𝜆2 𝐼 + 𝜆𝐷 0 + 𝛺 2 )−1 ∥∣𝜆∣ < 1,
𝐷′ from (19.11), which is insured if ∥(𝜆2 𝐼 + 𝜆𝐷 0 + 𝛺 2 )−1 ∥∥𝐷′ ∥∣𝜆∣ = Thus,
∥𝐷′ ∥∣𝜆∣ min𝑗 (∣𝜆 − 𝜆𝑗+ ∣∣𝜆 − 𝜆𝑗− ∣)
𝜎(𝐴) ⊆ ∪𝑗 𝒞(𝜆𝑗+ , 𝜆𝑗− , ∥𝐷 ′ ∥).
< 1.
(19.17)
19 Modal Approximation
151
These ovals will always have both foci either real √ or complex conjugate. If 𝑟 = ∥𝐷′ ∥ is small with respect to ∣𝜆𝑗+ − 𝜆𝑗− ∣ = ∣𝑑2𝑗𝑗 − 4𝜔𝑗2 ∣ then either ∣𝜆 − 𝜆𝑗+ ∣ or ∣𝜆− 𝜆𝑗− ∣ is small. In the first case the inequality ∣𝜆− 𝜆𝑗+ ∣∣𝜆 − 𝜆𝑗− ∣ ≤ ∣𝜆∣𝑟 is approximated by ⎧ √ 2𝜔𝑗 2 𝑑𝑗𝑗 < 2𝜔𝑗 𝑑𝑗𝑗 −4𝜔𝑗 𝑗 ⎨ ∣𝜆 ∣𝑟 ∣𝜆 − 𝜆𝑗+ ∣ ≤ 𝑗 + 𝑗 = 𝑟 √ ∣𝜆+ − 𝜆− ∣ − 𝑑2𝑗𝑗 −4𝜔𝑗2 ⎩ 𝑑𝑗𝑗√ 𝑑𝑗𝑗 > 2𝜔𝑗 2 2 𝑑𝑗𝑗 −4𝜔𝑗
and in the second
∣𝜆 − 𝜆𝑗− ∣ ≤
∣𝜆𝑗− ∣𝑟 ∣𝜆𝑗+ − 𝜆𝑗− ∣
⎧ √ 2𝜔𝑗 2 𝑑𝑗𝑗 < 2𝜔𝑗 ⎨ 𝑑𝑗𝑗 −4𝜔𝑗 =𝑟 . √ + 𝑑2𝑗𝑗 −4𝜔𝑗2 ⎩ 𝑑𝑗𝑗√ 𝑑𝑗𝑗 > 2𝜔𝑗 2 2 𝑑𝑗𝑗 −4𝜔𝑗
This is again a union of disks. If 𝑑𝑗𝑗 ≈ 0 then their radius is ≈ 𝑟/2. If 𝑑𝑗𝑗 ≈ 2𝜔𝑗 i.e. 𝜆− = 𝜆+ ≈ −𝑑𝑗𝑗 /2 the ovals look like a single circular disk. For large 𝑑𝑗𝑗 the oval around the absolutely larger eigenvalue is ≈ 𝑟 (the same behaviour as with (19.15)) whereas the smaller eigenvalue has the diameter ≈ 2𝑟𝜔𝑗2 /𝑑2𝑗𝑗 which is drastically better than (19.15). In the same way as before the Gershgorin type estimate is obtained 𝜎(𝐴) ⊆ ∪𝑗 𝒞(𝜆𝑗+ , 𝜆𝑗− , 𝑟𝑗 ).
(19.18)
We have called 𝐷0 a modal approximation to 𝐷 because the matrix 𝐷 is not uniquely determined by the input matrices 𝑀, 𝐶, 𝐾. Different choices of the transformation matrix 𝛷 give rise to different modal approximations 𝐷0 but the differences between them are mostly non-essential. To be more precise, let 𝛷 and 𝛷˜ both satisfy (2.4). Then 𝑀 = 𝛷−𝑇 𝛷−1 = 𝛷˜−𝑇 𝛷˜−1 , ˜−𝑇 𝛺 2 𝛷˜−1 𝐾 = 𝛷−𝑇 𝛺 2 𝛷−1 = 𝛷 ˜ is an orthogonal matrix which commutes with 𝛺 implies that 𝑈 = 𝛷−1 𝛷 which we now write in the form (by labelling 𝜔𝑗 differently) 𝛺 = diag(𝜔1 𝐼𝑛1 , . . . , 𝜔𝑠 𝐼𝑛𝑠 ),
𝜔1 < ⋅ ⋅ ⋅ < 𝜔𝑠 .
(19.19)
Denote by 𝐷 = (𝐷𝑖𝑗 ) the corresponding partition of the matrix 𝐷 and set 𝑈 0 = diag(𝑈11 , . . . , 𝑈𝑠𝑠 ),
152
19 Modal Approximation
where each 𝑈𝑗𝑗 is an orthogonal matrix of order 𝑛𝑗 from (19.10). Now, ˜ := 𝛷˜𝑇 𝐶 𝛷 ˜ = 𝑈 𝑇 𝛷∗ 𝐶𝛷𝑈 = 𝑈 𝑇 𝐷𝑈, 𝐷 ˜ 𝑖𝑗 = 𝑈𝑖𝑖𝑇 𝐷𝑖𝑗 𝑈𝑗𝑗 𝐷 and hence
˜ ′ = 𝑈 𝑇 𝐷′ 𝑈. 𝐷
Now, if the undamped frequencies are all simple, then 𝑈 is diagonal and the estimates (19.3) or (19.5) and (19.6) remain unaffected by this change of coordinates. Otherwise we replace diag(𝑑11 , . . . , 𝑑𝑛𝑛 ) by 𝐷0 = diag(𝐷11 , . . . , 𝐷𝑠𝑠 ),
(19.20)
where 𝐷0 commutes with 𝛺. In fact, a general definition of any modal approximation is that it 1. Is block-diagonal part of 𝐷 and 2. Commutes with 𝛺. The modal approximation with the coarsest possible partition – this is the one whose block dimensions equal the multiplicities in 𝛺 – is called a maximal modal approximation. Accordingly, we say that 𝐶 0 = 𝛷−1 𝐷0 𝛷−𝑇 is a modal approximation to 𝐶 (and also 𝑀, 𝐶 0 , 𝐾 to 𝑀, 𝐶, 𝐾). Proposition 19.5 Each modal approximation to 𝐶 is of the form 𝐶0 =
𝑠 ∑
𝑃𝑘∗ 𝐶𝑃𝑘 ,
𝑘=1
where 𝑃1 , . . . .𝑃𝑠 is an 𝑀 -orthogonal decomposition of the identity (that is 𝑃𝑘∗ = 𝑀 𝑃𝑘 𝑀 −1 ) and 𝑃𝑘 commute with the matrix √
𝑀 −1 𝐾 = 𝑀 −1/2
Proof. Use the formula 𝐷0 =
√
𝑀 −1/2 𝐾𝑀 −1/2 𝑀 1/2 .
𝑠 ∑
𝑃𝑘0 𝐷𝑃𝑘0
𝑘=1
with 𝑃𝑘0 = diag(0, . . . , 𝐼𝑛𝑘 , . . . , 0), and set 𝑃𝑘 = 𝛷𝑃𝑘0 𝛷−1 . Q.E.D.
𝐷 = 𝛷𝑇 𝐶𝛷,
𝐷0 = 𝛷𝑇 𝐶 0 𝛷
19 Modal Approximation
153
Exercise 19.6 Characterise the set of all maximal modal approximations, again in terms of the original matrices 𝑀, 𝐶, 𝐾. It is obvious that the maximal approximation is the best among all modal approximations in the sense that
where
ˆ 0 ∥𝐸 , ∥𝐷 − 𝐷 0 ∥𝐸 ≤ ∥𝐷 − 𝐷
(19.21)
ˆ 11 , . . . , 𝐷 ˆ 𝑧𝑧 ) ˆ 0 = diag(𝐷 𝐷
(19.22)
ˆ 𝑖𝑗 ) is any block partition of 𝐷 which is finer than that in (19.20). and 𝐷 = (𝐷 It is also obvious that 𝐷0 is better than any diagonal damping matrix e.g. the one given by (19.9). We will now prove that the inequality (19.21) is valid for the spectral norm also. We shall need the following Proposition 19.7 Let 𝐻 = (𝐻𝑖𝑗 ) be any partitioned Hermitian matrix such that the diagonal blocks 𝐻𝑖𝑖 are square. Set 𝐻 0 = diag(𝐻11 , . . . , 𝐻𝑠𝑠 ), Then
𝐻 ′ = 𝐻 − 𝐻 0.
𝜆𝑘 (𝐻) − 𝜆𝑛 (𝐻) ≤ 𝜆𝑘 (𝐻 ′ ) ≤ 𝜆𝑘 (𝐻) − 𝜆1 (𝐻),
(19.23)
where 𝜆𝑘 (⋅) denotes the non-decreasing sequence of the eigenvalues of any Hermitian matrix. Proof. The estimate (2.14) yields 𝜆𝑘 (𝐻) − max max 𝜎(𝐻𝑗𝑗 ) ≤ 𝜆𝑘 (𝐻 ′ ) ≤ 𝜆𝑘 (𝐻) − min min 𝜎(𝐻𝑗𝑗 ). 𝑗
𝑗
By the interlacing property, 𝜆1 (𝐻) ≤ 𝜎(𝐻𝑗𝑗 ) ≤ 𝜆𝑛 (𝐻). Altogether we obtain (19.23). Q.E.D. From (19.23) some simpler estimates immediately follow: ∥𝐻 ′ ∥ ≤ 𝜆𝑛 (𝐻) − 𝜆1 (𝐻) =: spread(𝐻)
(19.24)
and, if 𝐻 is positive (or negative) semidefinite ∥𝐻 ′ ∥ ≤ ∥𝐻∥.
(19.25)
154
19 Modal Approximation
Hence, in addition to (19.21) we also have ˆ 0 ∥. ∥𝐷 − 𝐷 0 ∥ ≤ ∥𝐷 − 𝐷 So, a best bound in (19.17) is obtained, if 𝐷 0 = 𝐷 − 𝐷′ is a maximal modal approximation. If 𝐷0 is block-diagonal and the corresponding 𝐷′ = 𝐷 − 𝐷0 is inserted in (19.17) then the values 𝑑𝑗𝑗 from (19.13) should be replaced by the corresponding eigenvalues of the diagonal blocks 𝐷𝑗𝑗 . But in this case we can further transform 𝛺 and 𝐷 by a unitary similarity 𝑈 = diag(𝑈1 , . . . , 𝑈𝑠 ) such that each of the blocks 𝐷𝑗𝑗 becomes diagonal (𝛺 stays unchanged). With this stipulation we may retain the formula (19.17) unaltered. This shows that taking just the diagonal part 𝐷0 of 𝐷 covers, in fact, all possible modal approximations, when 𝛷 varies over all matrices performing (2.4). Similar extension can be made to the bound (19.18) but then no improvements in general can be guaranteed although they are more likely to occur than not. By the usual continuity argument it is seen that the number of eigenvalues in each component of ∪𝑖 𝒞(𝜆𝑖+ , 𝜆𝑖− , 𝑟𝑖 ) is twice the number of involved diagonals. In particular, if we have the maximal number of 2𝑛 components, then each of them contains exactly one eigenvalue (Fig. 19.1). A strengthening in the sense of Brauer [8] is possible here also. We will show that the spectrum is contained in the union of double ovals, defined as 𝒟(𝜆𝑝+ , 𝜆𝑝− , 𝜆𝑞+ , 𝜆𝑞− , 𝑟𝑝 𝑟𝑞 ) = {𝜆 : ∣𝜆 − 𝜆𝑝+ ∣∣𝜆 − 𝜆𝑝− ∣∣𝜆 − 𝜆𝑞+ ∣∣𝜆 − 𝜆𝑞− ∣ ≤ 𝑟𝑝 𝑟𝑞 ∣𝜆∣2 }, where the union is taken over all pairs 𝑝 ∕= 𝑞 and 𝜆𝑝± are the solutions of 𝜆2 + 𝑑𝑝𝑝 𝜆 + 𝜔𝑝2 = 0 and similarly for 𝜆𝑞± . The proof just mimics the standard Brauer’s one. The quadratic eigenvalue problem is written as (𝜆2 + 𝜆𝑑𝑝𝑝 + 𝜔𝑝2 )𝑥𝑝 = −𝜆
𝑛 ∑
𝑑𝑝𝑗 𝑥𝑗 ,
(19.26)
𝑑𝑞𝑗 𝑥𝑗 ,
(19.27)
𝑗=1 𝑗∕=𝑝
(𝜆2 + 𝜆𝑑𝑞𝑞 + 𝜔𝑞2 )𝑥𝑞 = −𝜆
𝑛 ∑ 𝑗=1 𝑗∕=𝑞
where ∣𝑥𝑝 ∣ ≥ ∣𝑥𝑞 ∣ are the two absolutely largest components of 𝑥. If 𝑥𝑞 = 0 then 𝑥𝑗 = 0 for all 𝑗 ∕= 𝑝 and trivially 𝜆 ∈ 𝒟(𝜆𝑝+ , 𝜆𝑝− , 𝜆𝑞+ , 𝜆𝑞− , 𝑟𝑝 𝑟𝑞 ). If 𝑥𝑞 ∕= 0 then multiplying the equalities (19.26) and (19.27) yields
19 Modal Approximation
155
1.5
1.5
1
1
0.5
0.5
0
0
−0.5
−0.5
−1
−1
−1.5 −0.22 −0.2 −0.18 −0.16 −0.14 −0.12 −0.1 −0.08 −0.06 −0.04 −0.02
−1.5
−1 −0.9 −0.8 −0.7 −0.6 −0.5 −0.4 −0.3 −0.2 −0.1
0
Fig. 19.1 Ovals for 𝜔 = 1; 𝑑 = 0.1, 1; 𝑟 = 0.3
∣𝜆 − 𝜆𝑝+ ∣∣𝜆 − 𝜆𝑝− ∣∣𝜆 − 𝜆𝑞+ ∣∣𝜆 − 𝜆𝑞− ∣∣𝑥𝑝 ∣∣𝑥𝑞 ∣ ≤ ∣𝜆∣2
𝑛 𝑛 ∑ ∑
∣𝑑𝑝𝑗 ∣∣𝑑𝑞𝑘 ∣∣𝑥𝑗 ∣∣𝑥𝑘 ∣.
𝑗=1 𝑘=1 𝑗∕=𝑝 𝑘∕=𝑞
Because in the double sum there is no term with 𝑗 = 𝑘 = 𝑝 we always have ∣𝑥𝑗 ∣∣𝑥𝑘 ∣ ≤ ∣𝑥𝑝 ∣∣𝑥𝑞 ∣, hence the said sum is bounded by ∣𝜆∣2 ∣𝑥𝑝 ∣∣𝑥𝑞 ∣
𝑛 ∑ 𝑗=1 𝑗∕=𝑝
∣𝑑𝑝𝑗 ∣
𝑛 ∑
∣𝑑𝑞𝑘 ∣.
𝑘=1 𝑘∕=𝑞
Thus, our inclusion is proved. As it is immediately seen, the union of all double ovals is contained in the union of all stretched Cassini ovals (Fig. 19.2). The simplicity of the modal approximation suggests to try to extend it to as many systems as possible. A close candidate for such extension is any system with tightly clustered undamped frequencies, that is, 𝛺 is close to an 𝛺 0 from (19.19). Starting again with
156
19 Modal Approximation
1.5
1.5
1
1
0.5
0.5
0
0
−0.5
−0.5
−1
−1
−1.5
−1.6 −1.4 −1.2 −1
−0.8 −0.6 −0.4 −0.2
0
−1.5
−2
−1.5
1.5
1
0.5
0
−0.5
−1
−1.5
−2
−1.5
−1
−0.5
Fig. 19.2 Ovals for 𝜔 = 1; 𝑑 = 1.7, 2.3, 2.2; 𝑟 = 0.3, 0.3, 0.1
0
−1
−0.5
0
19 Modal Approximation
157
(𝜆2 𝐼 + 𝜆𝐷 + 𝛺 2 )−1 = (𝜆2 𝐼 + 𝜆𝐷0 + (𝛺 0 )2 )−1 (𝐼 + (𝜆𝐷′ + 𝑍) + (𝜆2 𝐼 + 𝜆𝐷 0 + (𝛺 0 )2 )−1 )−1 with 𝑍 = 𝛺 2 − (𝛺 0 )2 we immediately obtain ˆ 𝑗 , 𝜆𝑗 , ∥𝐷′ ∥, ∥𝑍∥), 𝜎(𝐴) ⊆ ∪𝑗 𝒞(𝜆 + −
(19.28)
where the set ˆ + , 𝜆− , 𝑟, 𝑞) = {𝜆 : ∣𝜆 − 𝜆+ ∣∣𝜆 − 𝜆− ∣ ≤ ∣𝜆∣𝑟 + 𝑞} 𝒞(𝜆 will be called modified Cassini ovals with foci 𝜆± and extensions 𝑟, 𝑞. Remark 19.8 The basis of any modal approximation is the diagonalisation of the matrix pair 𝑀, 𝐾. Now, an analogous procedure with similar results can be performed by diagonalising the pair 𝑀, 𝐶 or 𝐶, 𝐾. The latter would be recommendable in Example 1.8 because there the corresponding matrices ℛ and 𝒦 are already diagonal. There is a third possibility: to approximate the system 𝑀, 𝐶, 𝐾 by some 𝑀0 , 𝐶0 , 𝐾0 which is modally damped. Here we change all three system matrices and expect to get closer to the original system and therefore to obtain better bounds. For an approach along these lines see [45]. Exercise 19.9 Try to apply the techniques of this chapter to give bounds to the harmonic response considered in Chap. 18.
Chapter 20
Modal Approximation and Overdampedness
If the systems in the previous chapter are all overdamped then estimates are greatly simplified as complex regions become just intervals. But before going into this a more elementary – and more important – question arises: Can the modal approximation help to decide the overdampedness of a given system? After giving an answer to the latter question we will turn to the perturbation of the overdamped eigenvalues themselves. We begin with some obvious facts the proofs of which are left to the reader. Proposition 20.1 If the system 𝑀, 𝐶, 𝐾 is overdamped, then the same is true of the projected system 𝑀 ′ = 𝑋 ∗ 𝑀 𝑋,
𝐶 ′ = 𝑋 ∗ 𝐶𝑋,
𝐾 ′ = 𝑋 ∗ 𝐾𝑋,
where 𝑋 is any injective matrix. Moreover, the definiteness interval of the former is contained in the one of the latter. Proposition 20.2 Let 𝑀 = diag(𝑀11 , . . . , 𝑀𝑠𝑠 ) 𝐶 = diag(𝐶11 , . . . , 𝐶𝑠𝑠 ) 𝐾 = diag(𝐾11 , . . . , 𝐾𝑠𝑠 ). Then the system 𝑀, 𝐶, 𝐾 is overdamped, if and only if each of the systems 𝑀𝑗𝑗 , 𝐶𝑗𝑗 , 𝐾𝑗𝑗 is overdamped and their definiteness intervals have a non trivial intersection (which is then the definiteness interval of 𝑀, 𝐶, 𝐾) Corollary 20.3 If the system 𝑀, 𝐶, 𝐾 is overdamped, then the same is true of any of its modal approximations. Exercise 20.4 If a maximal modal approximation is overdamped, then so are all others. K. Veseli´ c, Damped Oscillations of Linear Systems, Lecture Notes in Mathematics 2023, DOI 10.1007/978-3-642-21335-9 20, © Springer-Verlag Berlin Heidelberg 2011
159
160
20 Modal Approximation and Overdampedness
In the following we shall need some well known sufficient conditions for negative definiteness of a general Hermitian matrix 𝐴 = (𝑎𝑖𝑗 ); these are: 𝑎𝑗𝑗 < 0 for all 𝑗 and either ∥𝐴 − diag(𝑎11 , . . . , 𝑎𝑛𝑛 )∥ < − max 𝑎𝑗𝑗 𝑗
(norm-diagonal dominance) or 𝑛 ∑
∣𝑎𝑘𝑗 ∣ < −𝑎𝑗𝑗 for all 𝑗
𝑘=1 𝑘∕=𝑗
(Gershgorin-diagonal dominance). Theorem 20.5 Let 𝛺, 𝐷, 𝑟𝑗 be from (2.5), (2.22), (19.16), respectively and 𝐷0 = diag(𝑑11 , . . . , 𝑑𝑛𝑛 ), Let
𝐷′ = 𝐷 − 𝐷0 .
𝛥𝑗 = (𝑑𝑗𝑗 − ∥𝐷′ ∥)2 − 4𝜔𝑗2 > 0 for all 𝑗
(20.1)
and −𝑑𝑗𝑗 + ∥𝐷′ ∥ − 𝑞− := max 𝑗 2 or
√
𝛥𝑗
−𝑑𝑗𝑗 + ∥𝐷′ ∥ + < min 𝑗 2
√
𝛥𝑗
=: 𝑞+ (20.2)
𝛥ˆ𝑗 = (𝑑𝑗𝑗 − 𝑟𝑗 )2 − 4𝜔𝑗2 > 0 for all 𝑗
(20.3)
and 𝑞ˆ− := max 𝑗
−𝑑𝑗𝑗 + 𝑟𝑗 − 2
√ 𝛥ˆ𝑗
< min
−𝑑𝑗𝑗 + 𝑟𝑗 +
𝑗
√ 𝛥ˆ𝑗
2
=: 𝑞ˆ+ .
(20.4)
Then the system 𝑀, 𝐶, 𝐾 is overdamped. Moreover, the interval (𝑞− , 𝑞+ ), (ˆ 𝑞− , 𝑞ˆ+ ), respectively, is contained in the definiteness interval of 𝑀, 𝐶, 𝐾. Proof. Let 𝑞− < 𝜇 < 𝑞+ . The negative definiteness of 𝜇2 𝐼 + 𝜇𝐷 + 𝛺 2 = 𝜇2 𝐼 + 𝜇𝐷0 + 𝛺 2 + 𝜇𝐷′ will be insured by norm-diagonal dominance, if −𝜇∥𝐷′ ∥ < −𝜇2 − 𝜇𝑑𝑗𝑗 − 𝜔𝑗2 for all 𝑗,
20 Modal Approximation and Overdampedness
161
that is, if 𝜇 lies between the roots of the quadratic equation 𝜇2 + 𝜇(𝑑𝑗𝑗 − ∥𝐷′ ∥) + 𝜔𝑗2 = 0 for all 𝑗 and this is insured by (20.1) and (20.2). The conditions (20.3) and (20.4) are treated analogously. Q.E.D. We are now prepared to adapt the spectral inclusion bounds from the previous chapter to overdamped systems. Recall that in this case the definiteness interval divides the 2𝑛 eigenvalues into two groups: 𝐽-negative and 𝐽-positive. Theorem 20.6 If (20.1) and (20.2) hold then the 𝐽-negative/𝐽-positive eigenvalues of 𝑄(⋅) are contained in ∪𝑗 (𝜇𝑗−− , 𝜇𝑗−+ ),
∪𝑗 (𝜇𝑗+− , 𝜇𝑗++ ),
respectively, with 𝜇𝑗++ =
−𝑑𝑗𝑗 − ∥𝐷′ ∥ ±
(𝑑𝑗𝑗 + ∥𝐷′ ∥)2 − 4𝜔𝑗2
(20.5)
2
−−
𝜇𝑗+− =
√
−𝑑𝑗𝑗 + ∥𝐷 ′ ∥ ±
√ (𝑑𝑗𝑗 − ∥𝐷′ ∥)2 − 4𝜔𝑗2 2
−+
.
(20.6)
An analogous statement holds, if (20.3) and (20.4) hold and 𝜇𝑗++ , 𝜇𝑗+− is −−
−+
replaced by 𝜇 ˆ𝑗++ , 𝜇 ˆ𝑗+− where in (20.5) and (20.6) ∥𝐷′ ∥ is replaced by 𝑟𝑗 . −−
−+
Proof. First note that the function 0 > 𝑧 → 𝑧 +
√
𝑧 2 − 4𝜔𝑗2 is decreasing, so
𝜇𝑗+− < 𝜇𝑗++ and similarly 𝜇𝑗−− < 𝜇𝑗−+ . Take first 𝑟 = ∥𝐷′ ∥. All spectra are real and negative, so we have to find the intersection of 𝒞(𝜆𝑗+ , 𝜆𝑗− , 𝑟) with the real line the foci 𝜆𝑗+ , 𝜆𝑗− from (19.13) being also real. This intersection will be a union of two intervals. For 𝜆 < 𝜆𝑗− and also for 𝜆 > 𝜆𝑗+ the 𝑗th ovals are given by (𝜆𝑗− − 𝜆)(𝜆𝑗+ − 𝜆) ≤ −𝜆𝑟 i.e.
𝜆2 − (𝜆𝑗+ + 𝜆𝑗− − 𝑟)𝜆 + 𝜆𝑗+ 𝜆𝑗− ≤ 0,
where 𝜆𝑗+ +𝜆𝑗− = −𝑑𝑗𝑗 and 𝜆𝑗+ 𝜆𝑗− = 𝜔𝑗2 . Thus, the left and the right boundary point of the real ovals are 𝜇𝑗++ . −−
162
20 Modal Approximation and Overdampedness
For 𝜆𝑗− < 𝜆 < 𝜆𝑗+ the ovals will not contain 𝜆, if (𝜆 − 𝜆𝑗− )(𝜆𝑗+ − 𝜆) ≤ −𝜆𝑟 i.e.
𝜆2 + (𝑑𝑗𝑗 − 𝑟)𝜆 + 𝜔𝑗2 < 0
with the solution
𝜇𝑗−+ < 𝜆 < 𝜇𝑗+− .
The same argument goes with 𝑟 = 𝑟𝑗 . Q.E.D. Note the inequality (𝜇𝑗−− , 𝜇𝑗−+ ) < (𝜇𝑘+− , 𝜇𝑘++ ) for all 𝑗, 𝑘. The results obtained in this chapter can be partially extended to nonoverdamped systems which have some 𝐽-definite eigenvalues. The idea is to use Theorem 12.15. Theorem 20.7 Suppose that there is a connected component 𝒞0 of the ovals 2 2 (19.17) containing only foci 𝜆+ 𝑗 with 𝑑𝑗𝑗 ≥ 4𝜔𝑗 . Then the eigenvalues of 𝑄(⋅) contained in 𝒞0 are all 𝐽-positive and their number (with multiplicities) equals the number of the foci in 𝒞0 . For these eigenvalues the estimates of Theorem 20.6 hold (the same for 𝐽-negatives). Proof. Let 𝐴 be from (3.10) and 𝐷 = 𝐷0 + 𝐷 ′ as in (19.10), [ [ ] ] 0 𝛺 0 𝛺 0 𝐴= , 𝐴 = . −𝛺 −𝐷0 −𝛺 −𝐷 ˆ = 𝐴(𝑡)
[
] 0 𝛺 , −𝛺 −𝐷 0 − 𝑡𝐷 ′
0 ≤ 𝑡 ≤ 1,
where 𝐷 (and therefore 𝐷 0 + 𝑡𝐷′ ) is positive semidefinite. The set of all ˆ satisfies the requirements in Theorem 12.15. In particular, by the such 𝐴(𝑡) ˆ are contained in the ones of 𝐴 and by property (19.7) the ovals of each 𝐴(𝑡) taking a contour 𝛤 which separates 𝒞0 from the rest of the ovals in (19.17) ˆ the relation 𝛤 ∩ 𝜎(𝐴(𝑡)) will hold and Theorem 12.15 applies. Now the ovals from 𝒞0 become intervals and Theorem 20.6 applies as well. Q.E.D.
20.1
Monotonicity-Based Bounds
As it is known for symmetric matrices monotonicity-based bounds for the eigenvalues ((2.14), (2.15) for a single matrix or (2.12) for a matrix pair) have an important advantage over Gershgorin-based bounds: While the latter are
20.1 Monotonicity-Based Bounds
163
merely inclusions, that is, the eigenvalue is contained in a union of intervals the former tell more: there each interval contains ‘its own eigenvalue’, even if it intersects other intervals. In this chapter we will derive bounds of this kind for overdamped systems. A basic fact is the following theorem Theorem 20.8 With overdamped systems the eigenvalues move asunder under growing viscosity. More precisely, let − + + 𝜆− 𝑛 ≤ ⋅ ⋅ ⋅ ≤ 𝜆1 < 𝜆1 ≤ ⋅ ⋅ ⋅ ≤ 𝜆𝑛 < 0
ˆ , 𝐶, ˆ 𝐾 ˆ is more be the eigenvalues of an overdamped system 𝑀, 𝐶, 𝐾. If 𝑀 ± ˆ viscous than the original one then its corresponding eigenvalues 𝜆𝑘 satisfy ˆ − ≤ 𝜆− , 𝜆 𝑘 𝑘
ˆ+ 𝜆+ 𝑘 ≤ 𝜆𝑘
A possible way to prove this theorem is to use Duffin’s minimax formulae [21]. Denoting the eigenvalues as in (10.3) the following formulae hold 𝜆+ 𝑘 = min max 𝑝+ (𝑥), 𝑆𝑘 𝑥∈𝑆𝑘
𝜆− 𝑘 = max min 𝑝− (𝑥), 𝑆𝑘 𝑥∈𝑆𝑘
where 𝑆𝑘 is any 𝑘-dimensional subspace and 𝑝± is defined in (14.11). Now the proof of Theorem 20.8 is immediate, if we observe that 𝑝ˆ+ (𝑥) ≥ 𝑝+ (𝑥),
𝑝ˆ− (𝑥) ≤ 𝑝− (𝑥)
ˆ , 𝐶, ˆ 𝐾). ˆ Note that in for any 𝑥 (ˆ 𝑝± is the functional (14.11) for the system 𝑀 this case the representation 𝑝+ (𝑥) =
−2𝑥∗ 𝐾𝑥 √ 𝑥∗ 𝐶𝑥 + 𝛥(𝑥)
is more convenient. As a natural relative bound for the system matrices we assume ∣𝑥∗ 𝛿𝑀 𝑥∣ ≤ 𝜖𝑥∗ 𝑀 𝑥, with
ˆ − 𝑀, 𝛿𝑀 = 𝑀
∣𝑥∗ 𝛿𝐶𝑥∣ ≤ 𝜖𝑥∗ 𝐶𝑥, 𝛿𝐶 = 𝐶ˆ − 𝐶,
∣𝑥∗ 𝛿𝐾𝑥∣ ≤ 𝜖𝑥∗ 𝐾𝑥,
ˆ − 𝐾, 𝛿𝐻 = 𝐾
𝜖 < 1.
We suppose that the system 𝑀, 𝐶, 𝐾 is overdamped and modally damped. As in Exercise 14.6 one sees that the overdampedness of the perturbed system ˆ , 𝐶, ˆ 𝐾 ˆ is insured, if 𝑀 𝜖
0. It is clear that each 𝜆 𝑘 − ˆ as well. analogous bound holds for 𝜆 𝑘 The proof of the minimax formulae which were basic for the needed monotonicity is rather tedious. There is another proof of Theorem 20.8 which uses the analytic perturbation theory. This approach is of independent interest and we will sketch it here. The basic fact from the analytic perturbation theory is this. Theorem 20.9 Let 𝜆0 be, say, a 𝐽-positive eigenvalue of 𝑄(⋅) and let 𝑚 be its multiplicity. Set 𝜆2 (𝑀 + 𝜖𝑀1 ) + 𝜆(𝐶 + 𝜖𝐶1 ), +𝐾 + 𝜖𝐾1 𝑀1 , 𝐶1 , 𝐾1 arbitrary Hermitian. Choose any 𝒰𝑟 = {𝜆 ∈ ℂ; ∣𝜆 − 𝜆0 ∣ < 𝑟)}
(20.7)
20.1 Monotonicity-Based Bounds
with
165
(𝜎(𝑄(⋅)) ∖ {𝜆0 }) ∩ 𝒰𝑟 = ∅.
Then there is a real neighbourhood 𝒪 of 𝜖 = 0 such that for 𝜖 ∈ 𝒪 the eigenvalues of 𝑄(⋅, 𝜖) within 𝒰𝑟 together with the corresponding eigenvectors can be represented as real analytic functions 𝜆1 (𝜖), . . . , 𝜆𝑚 (𝜖),
(20.8)
𝑥1 (𝜖), . . . , 𝑥𝑚 (𝜖),
(20.9)
respectively. All these functions can be analytically continued along the real axis, and their 𝐽-positivity is preserved until any of these eigenvalues meets another eigenvalue of different type. The proof of this theorem is not much simpler than that of the minimax formulae (see [27]). But it is plausible and easy to memorise: the eigenpairs are analytic as long as the eigenvalues do not get mixed. Once one stipulates this the monotonicity is easy to derive, we sketch the key step. Take any pair 𝜆, 𝑥 from (20.8), (20.9), respectively (for simplicity we suppress the subscript and the variable 𝜖). Then differentiate the identity 𝑄(𝜆, 𝜖)𝑥 = 0 with respect to 𝜖 and premultiply by 𝑥∗ thus obtaining 𝜆′ = −
𝑥∗ (𝜆2 𝑀1 + 𝜆𝐶1 + 𝐾1 )𝑥 2𝜆𝑥∗ (𝑀 + 𝜖𝑀1 )𝑥 + 𝑥∗ (𝐶 + 𝜖𝐶1 )𝑥
=−
𝑥∗ (𝜆2 𝑀1 + 𝜆𝐶1 + 𝐾1 )𝑥 √ , 𝛥
(20.10)
where 𝛥 = 𝛥(𝜖) = (𝑥∗ (𝐶 + 𝜖𝐶1 )𝑥)2 − 4𝑥∗ (𝑀 + 𝜖𝑀1 )𝑥𝑥∗ (𝐾 + 𝜖𝐾1 )𝑥 is positive by the assumed 𝐽-positivity. By choosing −𝑀1 , 𝐶1 , −𝐾1 as positive semidefinite, the right hand side of (20.10) is non-negative (note that 𝜆 is negative). So, 𝜆 grows with 𝜖. If the system is overdamped then the proof of the monotonicity in ˆ , 𝐶, ˆ 𝐾 ˆ is more viscous Theorem 20.8 is pretty straightforward. If the system 𝑀 than 𝑀, 𝐶, 𝐾 then by setting ˆ − 𝑀, 𝑀1 = 𝑀
𝐶1 = 𝐶ˆ − 𝐶,
ˆ − 𝐾, 𝐾1 = 𝐾
we obtain positive semidefinite matrices −𝑀1 , 𝐶1 , −𝐾1 . Now form 𝑄(𝜆, 𝜖) as in (20.7), apply Theorem 20.9 and the formula (20.10) and set 𝜖 = 0, 1.
166
20 Modal Approximation and Overdampedness
A further advantage of this approach is its generality: the monotonicity, holds ‘locally’ just for any eigenvalue of definite type, the pencil itself need not be overdamped. That is, with growing viscosity an eigenvalue of positive type moves to the right (and an eigenvalue ov the negative type moves to the left); this state of affairs lasts until an eigenvalue of different type is met. Exercise 20.10 Try to extend the criteria in Theorem 20.5 for the overdampedness to the case where some undamped frequencies are tightly clustered – in analogy to the bounds (19.28). Exercise 20.11 Try to figure out what happens, if in (20.10) the matrices −𝑀1 , 𝐶1 , −𝐾1 are positive semidefinite and the right hand side vanishes, say, for 𝜖 = 0. Exercise 20.12 Using (20.10) show that for small and growing damping the eigenvalues move into the left plane.
Chapter 21
Passive Control
Large oscillations may be dangerous to the vibrating system and are to be avoided in designing such systems. In many cases the stiffness and the mass matrix are determined by the static requirements so the damping remains as the regulating factor. How to tell whether a given system is well damped? Or, which quantity has to be optimised to obtain a best damped system within a given set of admissible systems (usually determined by one or several parameters)? These are the questions which we pose and partly solve in this chapter. A common criterion says: take the damping which produces the least spectral abscissa (meaning that its absolute value should be maximal). This criterion is based on the bounds (13.8) and (13.6). Consider once more the one-dimensional oscillator from Example 9.3. The minimal spectral abscissa is attained at the critical damping 𝑑 = 2𝜔 and its optimal value is 𝑑 𝜆𝑎𝑜 = − . 2 This value of 𝑑 gives a best asymptotic decay rate in (13.7) as well as a ‘best behaviour’ at finite times as far as it can be qualitatively seen on Fig. 13.1 in Chap. 13. However, in general, optimising the spectral abscissa gives no control for ∥𝑒𝐴𝑡 ∥ for all times. Another difficulty with the spectral abscissa is that it is not a smooth function of the elements of 𝐴 and its minimisation is tedious, if using common optimisation methods. In this chapter we will present a viable alternative curing both shortcomings mentioned above. We say: an optimal damping extracts as much energy as possible from the system. A possible rigorous quantification of this requirement is to ask ∫ 0
∞
𝑇
𝑦(𝑡) 𝐵𝑦(𝑡)𝑑𝑡 =
∫ 0
∞
𝑇
𝑦0𝑇 𝑒𝐴 𝑡 𝐵𝑒𝐴𝑡 𝑦0 𝑑𝑡 = 𝑦0𝑇 𝑋𝑦0 = min,
K. Veseli´ c, Damped Oscillations of Linear Systems, Lecture Notes in Mathematics 2023, DOI 10.1007/978-3-642-21335-9 21, © Springer-Verlag Berlin Heidelberg 2011
(21.1)
167
168
21 Passive Control
where 𝐵 is some positive semidefinite matrix serving as a weight and 𝑋 is the solution of the Lyapunov equation (13.9). This is the total energy, averaged over the whole time history of the free oscillating system.1 In order to get rid of the dependence on the initial data 𝑦0 we average once more over all 𝑦0 with the unit norm. We thus obtain the penalty function, that is, the function to be minimised ∫ ℰ = ℰ(𝐴, 𝜌) = 𝑦0𝑇 𝑋𝑦0 𝑑𝜌, ∥𝑦0 ∥=1
where 𝑑𝜌 is a given non-negative measure on the unit sphere in ℝ𝑛 .2 This measure, like the matrix 𝐵, has the role of a weight, provided by the user. Such measures can be nicely characterised as the following shows. Proposition 21.1 For any non-negative measure 𝑑𝜌 on the unit sphere in ℝ𝑛 there is a unique symmetric positive semidefinite matrix 𝑆 such that ∫ 𝑦0𝑇 𝑋𝑦0 𝑑𝜌 = Tr(𝑋𝑆) (21.2) ∥𝑦0 ∥=1
for any real symmetric matrix 𝑋. Proof. Consider the set of all symmetric matrices as a vector space with the scalar product ⟨𝑋, 𝑌 ⟩ = Tr(𝑋𝑌 ). The map
∫ 𝑋 →
∥𝑦0 ∥=1
𝑦0𝑇 𝑋𝑦0 𝑑𝜌
(21.3)
is obviously a linear functional and hence is represented as ⟨𝑋, 𝑆⟩ = Tr(𝑋𝑆), where 𝑆 is a uniquely determined symmetric matrix. To prove its positive semidefiniteness recall that by (21.2) the expression Tr(𝑋𝑆) is non-negative whenever 𝑋 is positive semidefinite. Suppose that 𝑆 has a negative eigenvalue 𝜆 and let 𝑃𝜆 be the projection onto the corresponding eigenspace. Now take 𝑋 = 𝑃𝜆 + 𝜖(𝐼 − 𝑃𝜆 ),
𝜖 > 0.
1 The
integral in (21.1) has the dimension of ‘action’ (energy × time) and so the principle (21.1) can be called a ‘minimum action principle’ akin to other similar principles in Mechanics. 2 In
spite of the fact that we will make some complex field manipulations with the matrix 𝐴 we will, until the rest of this chapter, take 𝐴 as real.
21 Passive Control
169
Then 𝑋 is positive definite and Tr(𝑋𝑆) = 𝜆𝑃𝜆 + 𝜖(𝐼 − 𝑃𝜆 )𝑆 would be negative for 𝜖 small enough – a contradiction. Q.E.D. Thus, we arrive at the following optimisation problem: find the minimum { } min Tr(𝑋𝑆) : 𝐴𝑇 𝑋 + 𝑋𝐴 = −𝐵 , (21.4) 𝐴∈𝒜
where 𝐵 and 𝑆 are given positive semidefinite real symmetric matrices and 𝒜 is the set of allowed matrices 𝐴. Typically 𝒜 will be determined by the set of allowed dampings 𝐶 in (3.3). A particularly distinguished choice is 𝑆 = 𝐼/(2𝑛), which correponds to the Lebesgue measure (this is the measure which ‘treats equally’ any vector from the unit sphere). Remark 21.2 The previous proposition does not quite settle the uniqueness of 𝑆 as we might want it here. The question is: do all asymptotically stable 𝐴 (or does some smaller relevant set of 𝐴-s over which we will optimise) uniquely define the functional (21.3)? We leave this question open. Exercise 21.3 Let a real 𝐴 be asymptotically stable and let (13.9) hold. Then Tr(𝑋𝑆) = Tr(𝑌 𝐵), where 𝑌 solves
𝐴𝑌 + 𝑌 𝐴𝑇 = −𝑆.
Exercise 21.4 For the system from Example 1.1 determine the matrix 𝐵 in (21.1) from the requirement 1. Minimise the average kinetic energy of the 𝑘th mass point. 2. Minimise the average potential energy of the 𝑘th spring. Minimising the trace can become quite expensive, if one needs to solve a series of Lyapunov equations in course of a minimisation process even on medium size matrices. In this connection we note an important economy in computing the gradient of the map 𝑋 → Tr(𝑆𝑋). We set 𝐴(𝛼) = 𝐴 + 𝛼𝐴1 , 𝐴1 arbitrary and differentiate the corresponding Lyapunov equation: 𝐴(𝛼)𝑇 𝑋 ′ (𝛼) + 𝑋 ′ (𝛼)𝐴(𝛼) = −𝐴′ (𝛼)𝑇 𝑋(𝛼) − 𝑋(𝛼)𝐴′ (𝛼) = −𝐴𝑇1 𝑋(𝛼) − 𝑋(𝛼)𝐴1 . Now, this is another Lyapunov equation of type (13.9), with 𝐴 = 𝐴(𝛼) and 𝐵 = 𝐴𝑇1 𝑋(𝛼) + 𝑋(𝛼)𝐴1 hence, using (13.11), Tr(𝑆𝑋(𝛼))′ =
∫ 0
∞
𝑇
Tr(𝑒𝐴𝑡 𝑆𝑒𝐴 𝑡 𝐴1 𝑋(𝛼))𝑑𝑡 +
∫ 0
∞
𝑇
Tr(𝑒𝐴𝑡 𝑆𝑒𝐴 𝑡 𝑋(𝛼)𝐴1 )𝑑𝑡
170
21 Passive Control
and the directional derivative along 𝐴1 reads 𝑑Tr(𝑆𝑋) = Tr(𝑆𝑋(𝛼))′𝛼=0 = Tr(𝑌 𝐴𝑇1 𝑋) + Tr(𝑋𝐴1 𝑌 ) = 2Tr(𝑋𝐴1 𝑌 ), 𝑑𝐴1 (21.5) where 𝑌 solves the equation 𝐴𝑌 + 𝑌 𝐴𝑇 = −𝑆. Thus, after solving just two Lyapunov equations we obtain all directional derivatives of the trace for free. A similar formula exists for the Hessian matrix as well.
21.1
More on Lyapunov Equations
The Lyapunov equation plays a central role in our optimisation approach. We will review here some of its properties and ways to its solution. As we have said the Lyapunov equation can be written as a system of 𝑛2 linear equations with as many unknowns. The corresponding linear operator ℒ = 𝐴𝑇 ⋅ + ⋅ 𝐴,
(21.6)
is known under the name of the Lyapunov operator and it maps the vector space of matrices of order 𝑛 into itself. Thus, Gaussian eliminations would take some (𝑛2 )3 = 𝑛6 operations which is feasible for matrices of very low order – not much more than 𝑛 = 50 on present day personal computers. The situation is different, if the matrix 𝐴 is safely diagonalised: 𝐴 = 𝑆𝛬𝑆 −1 ,
𝛬 = diag(𝜆1 , 𝜆2 , . . .).
Then the Lyapunov equation (13.9) is transformed into 𝛬𝑍 + 𝑍𝛬 = −𝐹,
𝑍 = 𝑆 𝑇 𝑋𝑆,
𝐹 = 𝑆 𝑇 𝐵𝑆
with the immediate solution 𝑍𝑖𝑗 =
−𝐹𝑖𝑗 . 𝜆𝑖 + 𝜆𝑗
The diagonalisation may be impossible or numerically unsafe if 𝑆 is illconditioned. A much more secure approach goes via the triangular form.
21.1 More on Lyapunov Equations
171
If the matrix 𝐴 is of, say, upper triangular form then the first row of (13.9) reads (𝑎11 + 𝑎11 )𝑥1𝑗 + 𝑎12 𝑥2𝑗 + ⋅ ⋅ ⋅ + 𝑎1𝑗 𝑥𝑗𝑗 = 𝑏1𝑗 from which the first row of 𝑋 is immediately computed; other rows are computed recursively. The solution with a lower triangular 𝑇 is analogous. A general matrix is first reduced to the upper triangular form as 𝑇 = 𝑈 ∗ 𝐴𝑈,
𝑈 unitary, possibly complex;
then (13.9) is transformed into 𝑇 ∗ 𝑌 + 𝑌 𝐴 = −𝑉,
𝑌 = 𝑈 ∗ 𝑋𝑈,
𝑉 = 𝑈 ∗ 𝐵𝑈,
which is then solved as above. (This is also another proof of the fact that the Lyapunov equation with an asymptotically stable 𝐴 possesses a unique solution.) Both steps to solve a general Lyapunov equation take some 25𝑛3 operations, where the most effort is spent to the triangular reduction. Thus far we have been optimising the transient behaviour only. It is important to know how this optimisation affects the steady state behaviour of the system. A key result to this effect is contained in the following Theorem 21.5 Let 𝐴 be real dissipative; then the integral 1 2𝜋𝑖
𝑌 =
∫
𝑖∞
−𝑖∞
(𝜆𝐼 − 𝐴𝑇 )−1 𝐵(𝜆𝐼 − 𝐴)−1 𝑑𝜆
(21.7)
exists, if 𝐴 is asymptotically stable. In this case 𝑌 equals 𝑋 from (13.9). Proof. For sufficiently large ∣𝜆∣ we have (
𝑇 −1
∥(𝜆𝐼 − 𝐴 )
−1
𝐵(𝜆𝐼 − 𝐴)
∞ ∑ ∥𝐴∥𝑘 ∥ ≤ ∥𝐵∥ ∣𝜆∣𝑘+1
)2 ,
𝑘=0
which is integrable at infinity. Thus, the integral in (21.7) exists, if there are no singularities on the imaginary axis, that is, if 𝐴 is asymptotically stable. In this case we have ∫ 𝑖∞ [ ] 1 𝐴𝑇 𝑌 = 𝜆(𝜆𝐼 − 𝐴𝑇 )−1 − 𝐼 𝐵(𝜆𝐼 − 𝐴)−1 𝑑𝜆, 2𝜋𝑖 −𝑖∞ 1 𝑌𝐴= 2𝜋𝑖
∫
𝑖∞
−𝑖∞
[ ] (𝜆𝐼 − 𝐴𝑇 )−1 𝐵 𝜆(𝜆𝐼 − 𝐴)−1 − 𝐼 𝑑𝜆.
By using the identities ∫ lim
𝜂→∞
𝑖𝜂
−𝑖𝜂
(𝜆𝐼 − 𝐴)−1 𝑑𝜆 = 𝑖𝜋𝐼,
∫ lim
𝜂→∞
𝑖𝜂 −𝑖𝜂
(𝜆𝐼 − 𝐴𝑇 )−1 𝑑𝜆 = 𝑖𝜋𝐼,
172
21 Passive Control
we obtain 𝐴𝑇 𝑌 + 𝑌 𝐴 = −𝐵 +
1 2𝜋𝑖
∫
𝑖∞
−𝑖∞
(𝜆 + 𝜆)(𝜆𝐼 − 𝐴𝑇 )−1 𝐵(𝜆𝐼 − 𝐴)−1 𝑑𝜆 = −𝐵.
Since the Lyapunov equation (13.9) has a unique solution it follows 𝑌 = 𝑋. Q.E.D. Hence
∫ Tr(𝑋𝑆) =
∫ =
∫
∥𝑦0 ∥=1
𝑑𝜌
0
∞
∥𝐵 1/2 𝑒𝐴𝑡 𝑦0 ∥2 𝑑𝑡 =
∥𝑦0 ∥=1
1 2𝜋
𝑦0𝑇 𝑋𝑦0 𝑑𝜌
∫
∫ 𝑑𝜌
∥𝑦0 ∥=1
∞
−∞
∥𝐵 1/2 (𝑖𝜔𝐼 − 𝐴)−1 𝑦0 ∥2 𝑑𝜔
This is an important connection between the ‘time-domain’ and the ‘frequency-domain’ optimisation: the energy average over time and all unit initial data of the free vibration is equal to some energy average of harmonic responses taken over all frequencies and all unit external forces. Exercise 21.6 Use the reduction to triangular form to show that a general Sylvester equation 𝐶𝑋 + 𝑋𝐴 = −𝐵 is uniquely solvable, if and only if 𝜎(𝐶) ∩ 𝜎(−𝐴) = ∅. Hint: reduce 𝐴 to the upper triangular form and 𝐶 to the lower triangular form. Exercise 21.7 Give the solution of the Lyapunov equation if a non-singular 𝑆 is known such that 𝑆 −1 𝐴𝑆 is block-diagonal. Are there any simplifications, if 𝑆 is known to be 𝐽-unitary and 𝐴 𝐽-Hermitian? Exercise 21.8 The triangular reduction needs complex arithmetic even if the matrix 𝐴 is real. Modify this reduction, as well as the subsequent recursive solution of the obtained Lyapunov equation with a block-triangular 𝑇 . Exercise 21.9 Let 𝐴 = −𝐼 + 𝑁 , 𝑁 𝑝−1 ∕= 0, 𝑁 𝑝 = 0. Then the solution of the Lyapunov equation (13.9) is given by the formula 2𝑝−1
𝑋=
∑
𝑛=0
21.2
1 2𝑘+1
𝑛 ( ) ∑ 𝑛 (𝑁 ∗ )𝑘 𝐵𝑁 𝑛−𝑘 . 𝑘 𝑘=0
Global Minimum of the Trace
The first question in assessing the trace criterion (21.4) is: if we let the damping vary just over all positive semidefinite matrices, where will the minimum be taken? The answer is: the minimum is taken at 𝐷 = 2𝛺 in the representation (3.10) that is, at the ‘modal critical damping’ – a very appealing result which speaks for the appropriateness of the trace as
21.2 Global Minimum of the Trace
173
a measure of stability. Note that without loss of generality we may restrict ourselves to the modal representation (3.10). As we know, the matrix 𝐴 from (3.10) is unitarily similar to the one from (3.3), the same then holds for the respective solutions 𝑋 of the Lyapunov equation 𝐴𝑇 𝑋 + 𝑋𝐴 = −𝐼,
(21.8)
so the trace of 𝑋 is the same in both cases. First of all, by the continuity property of the matrix eigenvalues it is clear that the set 𝒞𝑠 of all symmetric damping matrices with any fixed 𝑀 and 𝐾 and an asymptotically stable 𝐴 is open. We denote by 𝒟𝑠+ the connected component of 𝒞𝑠 containing the set of positive semidefinite matrices 𝐷. It is natural to seek the minimal trace within the set 𝒟𝑠+ as damping matrices. Theorem 21.10 Let 𝛺 be given and let in the modal representation (3.10) the matrix 𝐷 vary over the set 𝒟𝑠+ . Then the function 𝐷 → Tr(𝑋) where 𝑋 = 𝑋(𝐷) satisfies (21.8) has a unique minimum point 𝐷 = 𝐷0 = 2𝛺. The proof we will offer is not the simplest possible but it sheds some light on the geometry of asymptotically stable damped systems. The difficulty is that the trace is not generally a convex function which precludes the use of the usual uniqueness argument. Our proof will consist of several steps and end up with interpolating the trace function on the set of the stationary points by a function which is strictly convex and takes its minimum on the modal damping matrix, thus capturing the uniqueness. Lemma 21.11 If ∥𝐶∥2 ≥ 4∥𝐾∥∥𝑀 ∥ then max Re 𝜎(𝐴) ≥ −
2∥𝐾∥ . ∥𝐶∥
(21.9)
Proof. Let 𝑥 be a unit vector with 𝐶𝑥 = ∥𝐶∥𝑥. Then 𝑥𝑇 (𝜆2+ 𝑀 + 𝜆+ 𝐶 + 𝐾)𝑥 = 0 with 𝜆+ = −
2𝑥𝑇 𝐾𝑥 2∥𝐾∥ √ ≥− . 2 𝑇 𝑇 ∥𝐶∥ ∥𝐶∥ + ∥𝐶∥ − 4𝑥 𝐾𝑥𝑥 𝑀 𝑥
Since 𝜆2 𝑀 +𝜆𝐶 +𝐾 is positive definite for 𝜆 negative and close to zero, there must be an eigenvalue of 𝐴 in the interval [𝜆+ , 0) i.e. max Re 𝜎(𝐴) ≥ 𝜆+ . Q.E.D. The Lyapunov equation 𝐴𝑇 𝑋 + 𝑋𝐴 = −𝐼 for the phase-space matrix 𝐴 from (3.10) reads (𝐴0 − 𝐵𝐷𝐵 𝑇 )𝑇 𝑋 + 𝑋(𝐴0 − 𝐵𝐷𝐵 𝑇 ) = −𝐼,
(21.10)
174
where
21 Passive Control
[
] [ ] 0 𝛺 0 , 𝐵= . 𝐴0 = −𝛺 0 𝐼
The so-called dual Lyapunov equation is given by (𝐴0 − 𝐵𝐷𝐵 𝑇 )𝑌 + 𝑌 (𝐴0 − 𝐵𝐷𝐵 𝑇 )𝑇 = −𝐼.
(21.11)
To emphasise the dependence of 𝑋 and 𝑌 on 𝐷 we write 𝑋(𝐷) and 𝑌 (𝐷). By the 𝐽-symmetry of the matrix 𝐴 it follows 𝑌 (𝐷) = 𝐽𝑋(𝐷)𝐽.
(21.12)
Let us define the function 𝑓 : 𝒟𝑠+ → ℝ by 𝑓 (𝐷) = Tr(𝑋(𝐷)), where 𝑋(𝐷) solves (21.10).
(21.13)
Lemma 21.12 𝐷0 ∈ 𝒟𝑠+ is a stationary point of 𝑓 , if and only if 𝐵 𝑇 𝑋(𝐷0 )𝑌 (𝐷0 )𝐵 = 0.
(21.14)
Proof of the lemma. By setting in (21.5) 𝐴1 = 𝐵𝐷1 𝐵 𝑇 , where 𝐷1 is any real symmetric matrix the point 𝐷0 is stationary, if and only if Tr(𝑋(𝐷0 )𝐵𝐷1 𝐵 𝑇 𝑌 (𝐷0 )) = Tr(𝐷1 𝐵 𝑇 𝑌 (𝐷0 )𝑋(𝐷0 )𝐵) = 0. Now, by (21.12) the matrix 𝐵 𝑇 𝑌 (𝐷0 )𝑋(𝐷0 )𝐵 is symmetric: 𝐵 𝑇 𝑌 (𝐷0 )𝑋(𝐷0 )𝐵 = 𝐵 𝑇 𝐽𝑋(𝐷0 )𝐽𝑋(𝐷0 )𝐵, (𝐽𝑋(𝐷0 )𝐽𝑋(𝐷0 ))22 = −(𝑋(𝐷0 )𝐽𝑋(𝐷0 ))22 and (in sense of the trace scalar product) orthogonal to all symmetric matrices 𝐷1 , so it must vanish. Q.E.D. Proof of Theorem 21.10. By (13.12), Lemma 21.11 and the continuity of the map 𝐴 → 𝜆𝑎 (𝐴) the values of Tr𝑋 become infinite on the boundary of 𝒟𝑠+ . Hence there exists a minimiser and it is a stationary point. Let now 𝐷 be any such stationary point. By 𝑌 = 𝐽𝑋𝐽 (21.14) is equivalent to 2 𝑇 𝑋22 = 𝑋12 𝑋12 ,
(21.15)
21.2 Global Minimum of the Trace
where 𝑋 = 𝑋(𝐷) = We get
[
𝑋11 𝑋12 𝑇 𝑋12 𝑋22
175
] . Let us decompose (21.10) into components.
𝑇 −𝛺𝑋12 − 𝑋12 𝛺 = −𝐼,
(21.16)
𝑇 𝛺𝑋11 − 𝐷𝑋12 − 𝑋22 𝛺 = 0,
(21.17)
𝑇 𝛺𝑋12 + 𝑋12 𝛺 − 𝑋22 𝐷 − 𝐷𝑋22 = −𝐼.
(21.18)
From (21.16) it follows
1 (𝐼 − 𝑆)𝛺 −1 , (21.19) 2 where 𝑆 is skew-symmetric. From (21.17) and the previous formula it follows 𝑋12 =
1 −1 1 𝛺 𝐷𝛺 −1 + 𝛺 −1 𝐷𝛺 −1 𝑆 + 𝛺 −1 𝑋22 𝛺. 2 2
𝑋11 =
Since Tr(𝛺 −1 𝐷𝛺 −1 𝑆) = 0 we have Tr(𝑋) = Tr(𝑋11 ) + Tr(𝑋22 ) = From (21.15) it follows
1 Tr(𝛺 −1 𝐷𝛺 −1 ) + 2Tr(𝑋22 ). 2
𝑋12 = 𝑈 𝑋22 ,
(21.20) (21.21)
where 𝑈 is an orthogonal matrix. Note that the positive definiteness of 𝑋 implies that 𝑋22 is positive definite. Now (21.18) can be written as −𝑋22 𝐷 − 𝐷𝑋22 = −𝐼 − 𝛺𝑈 𝑋22 − 𝑋22 𝑈 𝑇 𝛺. Hence 𝐷 is the solution of a Lyapunov equation, and we have ∫ ∞ 𝐷= 𝑒−𝑋22 𝑡 (𝐼 + 𝛺𝑈 𝑋22 + 𝑋22 𝑈 𝑇 𝛺)𝑒−𝑋22 𝑡 d𝑡. 0
This implies 𝐷=
1 −1 𝑋 + 2 22
∫
∞ 0
𝑒−𝑋22 𝑡 𝛺𝑈 𝑒−𝑋22 𝑡 d𝑡 𝑋22 + 𝑋22
∫
∞
0
𝑒−𝑋22 𝑡 𝑈 𝑇 𝛺𝑒−𝑋22 𝑡 d𝑡.
From (21.19) and (21.21) follows 𝛺𝑈 =
) 1 ( −1 −1 𝑇 𝑋22 − 𝑋22 𝑈 𝑆𝑈 , 2
(21.22)
176
21 Passive Control
hence 𝐷=
−1 𝑋22
1 −1 − 𝑋22 2
∫
∞
0
𝑒−𝑋22 𝑡 𝑈 𝑇 𝑆𝑈 𝑒−𝑋22 𝑡 d𝑡 𝑋22 1 + 𝑋22 2
∫ 0
∞
−1 𝑒−𝑋22 𝑡 𝑈 𝑇 𝑆𝑈 𝑒−𝑋22 𝑡 d𝑡 𝑋22 .
So, we have obtained 1 −1 1 −1 −1 𝐷 = 𝑋22 − 𝑋22 𝑊 𝑋22 + 𝑋22 𝑊 𝑋22 , 2 2
(21.23)
where 𝑊 is the solution of the Lyapunov equation − 𝑋22 𝑊 − 𝑊 𝑋22 = −𝑈 𝑇 𝑆𝑈.
(21.24)
Since 𝑈 𝑇 𝑆𝑈 is skew-symmetric, so is 𝑊 . From (21.24) it follows −1 −1 𝑋22 𝑊 𝑋22 = −𝑊 + 𝑈 𝑇 𝑆𝑈 𝑋22 ,
which together with (21.23) implies 1 1 −1 𝑇 −1 −1 𝐷 = 𝑋22 + 𝑈 𝑇 𝑆𝑈 𝑋22 − 𝑋22 𝑈 𝑆𝑈. 2 2 Now taking into account relation (21.22) we obtain 𝐷 = 𝛺𝑈 + 𝑈 𝑇 𝛺, hence
𝛺 −1 𝐷𝛺 −1 = 𝑈 𝛺 −1 + 𝛺 −1 𝑈 𝑇 .
(21.25)
From (21.19) and (21.21) we obtain 𝑆 = 𝐼 − 2𝑈 𝑋22 𝛺,
(21.26)
and from this and the fact that 𝑆 is skew-symmetric it follows that 𝛺𝑋22 𝑈 𝑇 + 𝑈 𝑋22 𝛺 = 𝐼. This implies −1 −1 −1 −1 𝛺 −1 𝑈 𝑇 = 𝛺 −1 𝑋22 𝛺 − 𝛺 −1 𝑋22 𝛺 𝑈 𝑋22 𝛺.
(21.27)
21.2 Global Minimum of the Trace
177
From the previous relation it follows that −1 −1 Tr(𝛺 −1 𝑈 𝑇 ) = Tr(𝛺 −1 𝑋22 𝛺 ) − Tr(𝑈 𝛺 −1 ).
Now (21.20), (21.25) and the previous relation imply Tr(𝑋(𝐷)) = 𝑔(𝑋22 (𝐷)) with
1 Tr(𝛺 −1 𝑍 −1 𝛺 −1 ) + 2Tr(𝑍), 2 for any 𝐷 satisfying (21.14). Obviously, 𝑔 is positive-valued as defined on the set of all symmetric positive definite matrices 𝑍. We compute 𝑔 ′ , 𝑔 ′′ : 𝑔(𝑍) =
1 𝑔 ′ (𝑍)(𝑌ˆ ) = Tr(− 𝛺 −1 𝑍 −1 𝑌ˆ 𝑍 −1 𝛺 −1 + 2𝑌ˆ ) 2
(21.28)
𝑔 ′′ (𝑍)(𝑌ˆ , 𝑌ˆ ) = Tr(𝛺 −1 𝑍 −1 𝑌ˆ 𝑍 −1 𝑌ˆ 𝑍 −1 𝛺 −1 ) > 0 for any symmetric positive definite 𝑍 and any symmetric 𝑌ˆ ∕= 0. (The above expressions are given as linear/quadratic functionals; an explicit vector/matrix representation of the gradient and the Hessian would be clumsy whereas the functional representations are sufficient for our purposes.) By (21.28) the equation 𝑔 ′ (𝑍) = 0 means 1 Tr(− 𝛺 −1 𝑍 −1 𝑌ˆ 𝑍 −1 𝛺 −1 + 2𝑌ˆ ) = 0 2 for each symmetric 𝑌ˆ . This is equivalent to 1 − 𝑍 −1 𝛺 −2 𝑍 −1 + 2𝐼 = 0 2 i.e. 𝑍 = 𝑍0 = 12 𝛺 −1 and this is the unique minimiser with the minimum value 2Tr(𝛺 −1 ) (here, too, one easily sees that 𝑔(𝑍) tends to infinity as 𝑍 approaches the boundary of the set of positive definite symmetric matrices). Now, 12 𝛺 −1 = 𝑋22 (2𝛺) and 𝐷0 = 2𝛺 is readily seen to satisfy (21.14). Hence the value 2Tr(𝛺 −1 ) = 𝑔(𝑋22 (𝐷0 )) = Tr(𝑋(𝐷0 )) is strictly less than any other stationary value Tr(𝑋(𝐷)). Q.E.D. As we see the optimal damping is always positive definite; this is in contrast to the spectral abscissa criterion, where examples will be given in which the minimal spectral abscissa is actually taken at an indefinite damping matrix that is, outside of the physically allowed region.
178
21 Passive Control
Exercise 21.13 Prove that to the matrix 𝐷0 = 2𝛺 there corresponds the damping matrix √ −𝑇 𝑇 𝐶 = 𝐶0 = 2𝐿2 𝐿−1 (21.29) 2 𝐾𝐿2 𝐿2 . Exercise 21.14 Prove the formula ⎡3 𝑋(𝐷0 ) = ⎣
21.3
2
𝛺 −1 12 𝛺 −1
1 −1 1 −1 2𝛺 2𝛺
⎤ ⎦.
Lyapunov Trace vs. Spectral Abscissa
We will make some comparisons of the two optimality criteria: the spectral abscissa and the Lyapunov trace. It is expected that the first will give a better behaviour at 𝑡 = ∞ whereas the second will take into account finite times as well. In comparisons we will deal with the standard Lyapunov equation 𝐴𝑇 𝑋 + 𝑋𝐴 = −𝐼. With this 𝑋 (13.10) reads ∥𝑒𝐴𝑠 ∥ ≤ 𝜅(𝑋)𝑒−𝑡/∥𝑋∥ whereas (13.12) reads 2𝜆𝑎 ≤ −
1 . ∥𝑋∥
A simple comparison is made on the matrix [ ] [ ] −1 𝛼 𝐴𝑡 −𝑡 −1 𝛼𝑡 𝐴= with 𝑒 = 𝑒 . 0 −1 0 −1 Here 𝜆𝑎 (𝐴) = −1 independently of the parameter 𝛼 which clearly influences ∥𝑒𝐴𝑡 ∥. This is seen from the standard Lyapunov solution which reads [
] 1/2 𝛼/4 𝑋= . 𝛼/4 1/2 + 𝛼2 /4 If 𝐴 is a phase-space matrix of a damped system no such drastic differences between 𝜆𝑎 (𝐴) and the growth of ∥𝑒𝐴𝑡 ∥ seem to be known.3 3 This
seems to be an interesting open question: how far away can be the quantities 2𝜆𝑎 (𝐴) and −1/∥𝑋∥ on general damped systems.
21.3 Lyapunov Trace vs. Spectral Abscissa
179
0.5
0
−0.5
−1 1
1.5
2
2.5
3
3.5
4
Fig. 21.1 Spectral abscissa and Lyapunov trace (1D)
Example 21.15 For the matrix 𝐴 from Example 9.3 and 𝐵 = 𝐼 the Lyapunov equation reads [
0 −𝜔 𝜔 −𝑑
][
] [ ][ ] [ ] 𝑥11 𝑥12 𝑥11 𝑥12 0 𝜔 −1 0 + = . 𝑥21 𝑥22 𝑥21 𝑥22 −𝜔 −𝑑 0 −1
By solving four linear equations with four unknowns we obtain [
𝑥11 𝑥12 𝑋= 𝑥21 𝑥22
] =
[1
𝑑
+
𝑑 1 2𝜔 2 2𝜔 1 1 2𝜔 𝑑
] ,
where
2 𝑑 + 𝑑 2𝜔 2 has a unique minimum at the critical damping 𝑑 = 2𝜔. Tr(𝑋) =
The spectral abscissa and the Lyapunov trace as functions of the damping 𝑑 are displayed in Fig. 21.1 (an additive constant sees that both curves better fit to one coordinate frame). The minimum position is the same; the trace is smooth whereas the spectral abscissa has a cusp at the minimum. Example 21.16 Reconsider Example 13.13 by replacing the spectral abscissa by the Lyapunov trace as a function of the parameter 𝑐.
180
21 Passive Control
Concerning optimal spectral abscissae the following theorem holds Theorem 21.17 [23] Set
𝜏 (𝐶) = 𝜆𝑎 (𝐴),
where 𝐴 is from (3.3), 𝑀, 𝐾 are arbitrary but fixed whereas 𝐶 varies over the set of all real symmetric matrices of order 𝑛. Then 𝜏 (𝐶) attains its minimum which is equal to ( )1 det(𝐾) 2𝑛 𝜏0 = . det(𝑀 ) The minimiser is unique, if and only if the matrices 𝐾 and 𝑀 are proportional. The proof is based on a strikingly simple fact with an even simpler and very illuminating proof: Proposition 21.18 Let 𝒜 be any set of matrices of order 2𝑛 with the following properties 1. det(𝐴) is constant over 𝒜, 2. The spectrum of each 𝐴 ∈ 𝒜 is symmetric with respect to the real axis, counting multiplicities, 3. Re 𝜎(𝐴) < 0 for 𝐴 ∈ 𝒜. Denote by 𝒜0 the subset of 𝒜 consisting of all matrices whose spectrum consists of a single point. If 𝒜0 is not empty then min 𝜆𝑎 (𝐴)
𝐴∈𝒜
is attained on 𝒜0 and nowhere else. Proof. Let 𝜆1 , . . . , 𝜆2𝑛 be the eigenvalues of 𝐴. Then 0 < det(𝐴) =
2𝑛 ∏ 𝑘=1
Thus,
𝜆𝑘 =
2𝑛 ∏ 𝑘=1
∣𝜆𝑘 ∣ ≥
2𝑛 ∏
∣ Re(𝜆𝑘 )∣ ≥ (𝜆𝑎 (𝐴))
2𝑛
.
(21.30)
𝑘=1
1
𝜆𝑎 (𝐴) ≥ − (det(𝐴)) 2𝑛
(21.31)
By inspecting the chain of inequalities in (21.30) it is immediately seen that the equality in (21.31) implies the equalities in (21.30) and this is possible, if and only if 𝜆𝑘 = 𝜆𝑎 (𝐴), 𝑘 = 1, . . . , 2𝑛. Q.E.D. The set 𝒜, of all asymptotically stable phase-space matrices 𝐴 with fixed 𝑀, 𝐾 obviously fulfills the conditions of the preceding proposition. Indeed,
21.3 Lyapunov Trace vs. Spectral Abscissa
181
det(𝐴) = 𝜏02𝑛 , hence the only thing which remains to be proved is: for given 𝑀, 𝐾 find a symmetric damping matrix 𝐶 such that the spectrum of 𝐴 consists of a single point. This is a typical inverse spectral problem where one is asked to construct matrices from a given class having prescribed spectrum. This part of the proof is less elementary, it borrows from the inverse spectral theory and will be omitted. As an illustration we bring a simple inverse spectral problem: Example 21.19 Consider the system in Example 4.5. We will determine the constants 𝑚, 𝑘1 , 𝑘2 , 𝑐 by prescribing the eigenvalues 𝜆1 , 𝜆2 , 𝜆3 with Re(𝜆𝑘 ) < 0 and, say, 𝜆1 real. Since the eigenvalues of (14.2) do not change, if the equation is multiplied by a constant, we may assume that 𝑚 = 1. The remaining constants are determined from the identity [ det
] 𝜆2 + 𝑘1 + 𝑘2 −𝑘2 = 𝑐𝜆3 + 𝑘2 𝜆2 + 𝑐(𝑘1 + 𝑘2 )𝜆 + 𝑘1 𝑘2 −𝑘2 𝜆𝑐 + 𝑘2
( ) = 𝑐 𝜆3 − (𝜆1 + 𝜆2 + 𝜆3 )𝜆2 + (𝜆1 𝜆2 + 𝜆2 𝜆3 + 𝜆3 𝜆1 )𝜆 − 𝜆1 𝜆2 𝜆3 . The comparison of the coefficients yields 𝑘2 = −𝑐(𝜆1 + 𝜆2 + 𝜆3 ), 𝑘1 + 𝑘2 = 𝜆1 𝜆2 + 𝜆2 𝜆3 + 𝜆3 𝜆1 , 𝑘1 𝑘2 = −𝑐𝜆1 𝜆2 𝜆3 . This system is readily solved with the unique solution 𝑘1 = 𝑘2 =
𝜆1 𝜆2 𝜆3 , 𝜆1 + 𝜆2 + 𝜆3
𝜆21 (𝜆2 + 𝜆3 ) + 𝜆22 (𝜆1 + 𝜆3 ) + 𝜆23 (𝜆1 + 𝜆2 ) + 2𝜆1 𝜆2 𝜆3 , 𝜆1 + 𝜆2 + 𝜆3 𝑐=−
𝑘2 . 𝜆1 + 𝜆2 + 𝜆3
Here we can apply Proposition 21.18 by taking first, say, 𝜆1 = 𝜆2 = 𝜆3 = −1 thus obtaining 𝑘1 = 1/3, 𝑘2 = 8/3, 𝑐 = 8/9. Since 𝜆2 , 𝜆3 are either real or complex conjugate the values 𝑘1 , 𝑘2 , 𝑐 are all positive.
182
21 Passive Control
Fig. 21.2 Time histories
Now with the fixed values 𝑘1 , 𝑘2 as above we let 𝑐 > 0 be variable. The so obtained set 𝒜 of phase-space matrices from Example 4.5 again satisfies the conditions of Proposition 21.18 and 𝑐 = 8/9 yields the optimal spectral abscissa. For numerical comparison between the two criteria we will take an example from [23]. Take [ ] [ ] 10 𝑘1 0 𝑀= , 𝐾= 01 0 𝑘2 with the optimal-abscissa damping [ ] 3/4 1/4 1/2 1/2 1 4𝑘1 𝑘2 ±(𝑘1 − 𝑘2 )2 𝐶 = 1/2 . 1/2 1/2 1/4 3/4 1/2 ±(𝑘1 − 𝑘2 )2 4𝑘1 𝑘2 𝑘1 + 𝑘2 Here the different sign choices obviously lead to unitarily equivalent both damping and phase-space matrices, so we will take the first one. According to Theorem 21.10 our best damping in this case is 𝐶𝑑 =
[ √ ] 2 𝑘1 0 √ . 0 2 𝑘2
21.3 Lyapunov Trace vs. Spectral Abscissa
183
For comparing time histories it is convenient to observe the quantity 𝑇
∥𝑒𝐴𝑡 ∥2𝐸 = Tr(𝑒𝐴 𝑡 𝑒𝐴𝑡 ) because (1) it dominates the square of the spectral norm of 𝑒𝐴𝑡 and (2) it equals the average over all unit vectors 𝑦0 of the energy ∥𝑒𝐴𝑡𝑦0 ∥2 at each time 𝑡. These histories can be seen in Fig. 21.2. The dashed line corresponds to the optimal trace and the full one to the optimal abscissa. The values we have taken are 𝑘1 = 1, 𝑘2 = 4, 16, 33, 81 The optimal abscissa history is better at large times but this difference is almost invisible since both histories are then small. At 𝑘2 about 33 the matrix 𝐶 ceases to be positive semidefinite; the history for 𝑘2 = 81 is even partly increasing in time and this becomes more and more drastic with growing 𝑘2 . In such situations the theory developed in [23] does not give information on an optimal-abscissa within positive-semidefinite dampings.
Chapter 22
Perturbing Matrix Exponential
Modal approximation appears to be very appealing to treat matrix exponential by means of perturbation techniques. We will now derive some abstract perturbation bounds and apply them to phase-space matrices of damped systems. The starting point will be the formula which is derived from (3.4) 𝑒𝐵𝑡 − 𝑒𝐴𝑡 =
∫
𝑡 0
𝑒𝐵(𝑡−𝑠) (𝐵 − 𝐴)𝑒𝐴𝑠 𝑑𝑠 =
∫ 0
𝑡
𝑒𝐵𝑠 (𝐵 − 𝐴)𝑒𝐴(𝑡−𝑠) 𝑑𝑠.
(22.1)
A simplest bound is obtained, if the exponential decay of the type (13.8) is known for both matrices. So, let (13.8) hold for 𝐴 and 𝐵 with the constants 𝐹, 𝜈, 𝐹1 , 𝜈1 , respectively. Then by (22.1) we immediately obtain ∥𝑒𝐵𝑡 − 𝑒𝐴𝑡 ∥ ≤ ∥𝐵 − 𝐴∥
= ∥𝐵 − 𝐴∥
∫ 0
𝑡
𝑒−𝜈𝑠 𝑒−𝜈1 (𝑡−𝑠) 𝐹 𝐹1 𝑑𝑠
(22.2)
⎧ 𝑒−𝜈𝑡 −𝑒−𝜈1 𝑡 ⎨ 𝐹 𝐹1 𝜈1 −𝜈 , 𝜈1 ∕= 𝜈 ⎩
𝐹 𝐹1 𝑡𝑒−𝜈𝑡 , 𝜈1 = 𝜈.
If 𝐵 is only known to be dissipative (𝐹1 = 1, 𝜈1 = 0) then ∥𝑒𝐵𝑡 − 𝑒𝐴𝑡 ∥ ≤ ∥𝐵 − 𝐴∥
𝐹 𝐹 (1 − 𝑒−𝜈𝑡 ) ≤ ∥𝐵 − 𝐴∥ . 𝜈 𝜈
(22.3)
If 𝐴, too, is just dissipative (𝐹 = 1, 𝜈 = 0) then ∥𝑒𝐵𝑡 − 𝑒𝐴𝑡 ∥ ≤ ∥𝐵 − 𝐴∥𝑡.
(22.4)
The perturbation estimates obtained above are given in terms of general asymptotically stable matrices. When applying this to phase-space matrices 𝐴, 𝐵 we are interested to see how small will 𝐵 − 𝐴 be, if 𝑀, 𝐶, 𝐾 are subject K. Veseli´ c, Damped Oscillations of Linear Systems, Lecture Notes in Mathematics 2023, DOI 10.1007/978-3-642-21335-9 22, © Springer-Verlag Berlin Heidelberg 2011
185
186
22 Perturbing Matrix Exponential
to small changes. This is called ‘structured perturbation’ because it respects the structure of the set of phase-space matrices within which the perturbation takes place. An example of structured perturbation was given in Chap. 19 on modal approximations. If 𝐵 is obtained from 𝐴 by changing 𝐶 into 𝐶ˆ then as in (19.3) we obtain the following estimate ∥(𝐵 − 𝐴)∥ = max ∣ 𝑥
𝑥𝑇 (𝐶ˆ − 𝐶)𝑥 ∣. 𝑥𝑇 𝑀 𝑥
Exercise 22.1 Let 𝐶 be perturbed as ∣𝑥𝑇 (𝐶ˆ − 𝐶)𝑥∣ ≤ 𝜖𝑥𝑇 𝐶𝑥. Prove ∥𝐵 − 𝐴∥ ≤ 𝜖 max 𝑥
𝑥𝑇 𝐶𝑥 . 𝑥𝑇 𝑀 𝑥
The dependence of 𝐴 on 𝑀 and 𝐾 is more delicate. This is because 𝑀 and 𝐾 define the phase-space itself and by changing them we are changing the underlying geometry of the system. First of all, recall that 𝐴 is defined by 𝑀, 𝐶, 𝐾 only up to (jointly) unitary equivalence. So, it may happen that in order to obtain best estimates we will have to take phase-space realisations not enumerated in (3.10). We consider ˆ and 𝑀, 𝐶 remain unchanged, that is, the case in which 𝐾 is perturbed into 𝐾 ˆ = 𝐾 + 𝛿𝐾 = 𝐿 ˆ 1𝐿 ˆ 𝑇1 𝐾 with
√ ˆ 1 = 𝐿1 𝐼 + 𝑍, 𝐿
−𝑇 𝑍 = 𝐿−1 1 𝛿𝐾𝐿1 ,
where we have assumed ∥𝑍∥ ≤ 1. Now, ˆ 𝐿−1 2 ( 𝐿1
− 𝐿1 ) =
𝐿−1 2 𝐿1
) ∞ ( ∑ 1/2 𝑘 𝑍 𝑘 𝑘=1
=
) ∞ ( ∑ 1/2 −1 −𝑇 𝑘−1 𝐿2 𝛿𝐾𝐿−𝑇 = 𝐿−1 1 𝑍 2 𝛿𝐾𝐿1 𝑓 (𝑍) 𝑘 𝑘=1
with
{ 𝑓 (𝜁) =
(1+𝜁)1/2 −1 , 𝜁
𝜁= ∕ 0 0, 𝜁 = 0
and ∣𝜁∣ ≤ 1. By 𝑓 (−1) = 1,
𝑓 ′ (𝜁) = −
((1 + 𝜁)1/2 + 1)2 ≤0 2𝜁 2 (1 + 𝜁)1/2
22 Perturbing Matrix Exponential
187
and by the fact that 𝑍 is a real symmetric matrix we have ∥𝑓 (𝑍)∥ ≤ 1 and −1 −𝑇 ˆ ∥𝐵 − 𝐴∥ = ∥𝐿−1 2 (𝐿1 − 𝐿1 )∥ ≤ ∥𝐿2 𝛿𝐾𝐿1 ∥.
Thus (22.3) yields ∥𝑒𝐵𝑡 − 𝑒𝐴𝑡 ∥ ≤
𝐹 −1 ∥𝐿 𝛿𝐾𝐿−𝑇 1 ∥. 𝜈 2
(22.5)
ˆ 1 will not be the Cholesky factor, if 𝐿1 was such. If we had Note that here 𝐿 ˆ 1 to be the Cholesky factor then much less sharp estimate would insisted 𝐿 be obtained. We now derive a perturbation bound which will assume the dissipativity of both 𝐴 and 𝐵 but no further information on their exponential decay – there might be none, in fact. By 𝒯 (𝐴) we denote the set of all trajectories { } 𝑆 = 𝑥 = 𝑒𝐴𝑡 𝑥0 , 𝑡 ≥ 0 , for some vector 𝑥0 . Lemma 22.2 Let 𝐴, 𝐵 be arbitrary square matrices. Suppose that there exist trajectories 𝑆 ∈ 𝒯 (𝐴), 𝑇 ∈ 𝒯 (𝐵 ∗ ) and an 𝜀 > 0 such that for any 𝑦 ∈ 𝑆, 𝑥∈𝑇 ∣𝑥∗ (𝐵 − 𝐴)𝑦∣2 ≤ 𝜀2 Re(−𝑥∗ 𝐵𝑥) Re(−𝑦 ∗ 𝐴𝑦). (22.6) Then for all such 𝑥, 𝑦 ∣𝑥∗ (𝑒𝐵𝑡 − 𝑒𝐴𝑡 )𝑦∣ ≤
𝜀 ∥𝑥∥∥𝑦∥. 2
(Note that in (22.6) it is tacitly assumed that the factors on the right hand side are non-negative. Thus, we assume that 𝐴, 𝐵 are ‘dissipative along a trajectory’.) Proof. Using (22.1), (22.6) and the Cauchy-Schwarz inequality we obtain ∣𝑥∗ (𝑒𝐵𝑡 − 𝑒𝐴𝑡 )𝑦∣2 ∫ ≤ ≤ 𝜀2
∫ 𝑡√ 0
≤ 𝜀2
∫
𝑡 0
0
𝑡
∗
∣(𝑒𝐵 𝑠 𝑥)∗ (𝐵 − 𝐴)𝑒𝐴(𝑡−𝑠) 𝑦∣𝑑𝑠
2
Re(−𝑥∗ 𝑒𝐵𝑠 𝐵𝑒𝐵 ∗ 𝑠 𝑥) Re(−𝑦 ∗ 𝑒𝐴∗ (𝑡−𝑠) 𝐴𝑒𝐴(𝑡−𝑠) 𝑦)𝑑𝑠 ∗
Re(−𝑥∗ 𝑒𝐵𝑠 𝐵𝑒𝐵 𝑠 𝑥)𝑑𝑠
∫
𝑡 0
∗
Re(−𝑦 ∗ 𝑒𝐴 𝑠 𝐴𝑒𝐴𝑠 𝑦)𝑑𝑠
2
188
22 Perturbing Matrix Exponential
By partial integration we compute ∫ ℐ(𝐴, 𝑦, 𝑡) =
𝑡
0
∗
𝑅𝑒(−𝑦 ∗ 𝑒𝐴 𝑠 𝐴𝑒𝐴𝑠 𝑦)𝑑𝑠
𝑡 = −∥𝑒𝐴𝑠 𝑦∥2 − ℐ(𝐴, 𝑦, 𝑡), 0
) 1( ℐ(𝐴, 𝑦, 𝑡) = ∥𝑦∥2 − ∥𝑒𝐴𝑡 𝑦∥2 2 (note that ℐ(𝐴, 𝑦, 𝑡) = ℐ(𝐴∗ , 𝑦, 𝑡)). Obviously 0 ≤ ℐ(𝐴, 𝑦, 𝑡) ≤
1 ∥𝑦∥2 2
and ℐ(𝐴, 𝑦, 𝑡) increases with 𝑡. Thus, there exist limits ∥𝑒𝐴𝑡𝑦∥2 ↘ 𝑃 (𝐴, 𝑦), ℐ(𝐴, 𝑦, 𝑡) ↗ ℐ(𝐴, 𝑦, ∞) =
𝑡→∞
) 1 𝑦∥2 − 𝑃 (𝐴, 𝑦) , 2
with 0 ≤ ℐ(𝐴, 𝑦, ∞) ≤
𝑡→∞
1 ∥𝑦∥2 . 2
(and similarly for 𝐵). Altogether ∣𝑥∗ (𝑒𝐵𝑡 − 𝑒𝐴𝑡 )𝑦∣2 ≤
)( ) 𝜀2 ( ∥𝑥∥2 − 𝑃 (𝐵 ∗ , 𝑥) ∥𝑦∥2 − 𝑃 (𝐴, 𝑦) 4 ≤
𝜀2 ∥𝑥∥2 ∥𝑦∥2 . 4
Q.E.D. Corollary 22.3 Suppose that (22.6) holds for all 𝑦 from some 𝑆 ∈ 𝒯 (𝐴) and all 𝑥. Then ( ) 𝜀 ∥ 𝑒𝐵𝑡 𝑦 − 𝑒𝐴𝑡 𝑦 ∥ ≤ ∥𝑦∥. 2 Proposition 22.4 For any dissipative 𝐴 we have 𝑃 (𝐴, 𝑦) = 𝑦 ∗ 𝑃 (𝐴)𝑦, where the limit
∗
𝑃 (𝐴) = lim 𝑒𝐴 𝑡 𝑒𝐴𝑡 𝑡→∞
exists, is Hermitian and satisfies 0 ≤ 𝑦 ∗ 𝑃 (𝐴)𝑦 ≤ 𝑦 ∗ 𝑦 for all 𝑦.
(22.7)
22 Perturbing Matrix Exponential
189
Proof. The existence of the limit follows from the fact that any bounded, non decreasing sequence of real numbers – and also of Hermitian matrices – is convergent. All other statements are then obvious. Q.E.D. Theorem 22.5 If both 𝐴 and 𝐵 are dissipative and (22.6) holds for all 𝑥, 𝑦, then ( ) )( ) 𝜀2 ( ∣ 𝑥, (𝑒𝐵𝑡 − 𝑒𝐴𝑡 )𝑦 ∣2 ≤ ∥𝑥∥2 − 𝑥∗ 𝑃 (𝐵)𝑥 ∥𝑦∥2 − 𝑦 ∗ 𝑃 (𝐴)𝑦 . 4 In particular, ∥𝑒𝐵𝑡 − 𝑒𝐴𝑡 ∥ ≤
𝜀 . 2
(22.8)
Exercise 22.6 With 𝐴 dissipative prove that 𝑃 (𝐴) from (22.7) is an orthogonal projection and find its range. Corollary 22.7 If in Theorem 22.5 we have 𝜀 < 2 then the asymptotic stability of 𝐴 implies the same for the 𝐵 and vice versa. Proof. Just recall that the asymptotic stability follows, if ∥𝑒𝐴𝑡 ∥ < 1 for some 𝑡 > 0. Q.E.D. We now apply the key result of this chapter, contained in Theorem 22.5 to phase-space matrices 𝐴, 𝐵 having common masses and stiffnesses and different damping matrices 𝐶, 𝐶1 , respectively. As is immediately seen then the bound (22.6) is equivalent to ∣𝑥∗ (𝐶1 − 𝐶)𝑦∣2 ≤ 𝜀2 𝑥∗ 𝐶𝑥𝑦 ∗ 𝐶1 𝑦.
(22.9)
Example 22.8 Consider the damping matrix 𝐶 = 𝐶𝑖𝑛 from (1.4) and (1.5) (1) and let 𝐶1 be of the same type with parameters 𝑐𝑗 and let the two sets of parameters be close in the sense (1)
∣𝑐𝑗 − 𝑐𝑗 ∣ ≤ 𝜖
√ (1) 𝑐𝑗 𝑐𝑗 .
Then (everything is real) (1)
𝑥𝑇 (𝐶1 − 𝐶)𝑦 = (𝑐1 − 𝑐1 )𝑥1 𝑦1 +
𝑛 ∑ 𝑗=2
(1)
(1)
(𝑐𝑗 − 𝑐𝑗 )(𝑥𝑗 − 𝑥𝑗−1 )(𝑦𝑗 − 𝑦𝑗−1 )
+(𝑐𝑛+1 − 𝑐𝑛+1 )𝑥𝑛 𝑦𝑛
190
22 Perturbing Matrix Exponential
and, by the Cauchy-Schwartz inequality, ∣𝑥∗ (𝐶1 − 𝐶)𝑦∣ ⎡ ⎤ √ √ 𝑛 √ ∑ (1) (1) (1) ≤ 𝜖 ⎣ 𝑐1 𝑐1 ∣𝑥1 𝑦1 ∣ + 𝑐𝑗 𝑐𝑗 ∣𝑥𝑗 − 𝑥𝑗−1 ∣∣𝑦𝑗 − 𝑦𝑗−1 ∣ + 𝑐𝑛+1 𝑐𝑛+1 ∣𝑥𝑛 𝑦𝑛 ∣⎦ 𝑗=2
≤𝜖
√ 𝑥𝑇 𝐶1 𝑥𝑦 𝑇 𝐶𝑦.
This shows how natural the bound (22.6) is. Exercise 22.9 Do the previous example for the case of perturbed 𝐶𝑜𝑢𝑡 in the analogous way. Note the difference between (22.6), (22.8) and (22.2)–(22.4): the latter use the usual norm measure for 𝐵 − 𝐴 but need more information on the decay of 𝑒𝐴𝑡 whereas the former need no decay information at all (except, of course, the dissipativity for both), in fact, neither 𝐴 nor 𝐵 need be asymptotically stable. On the other hand, the right hand side of (22.2) goes to zero as 𝑡 → 0 while the right hand side of (22.4) diverges with 𝑡 → ∞. Exercise 22.10 How does the bound (22.9) affect the phase-space matrices of the type (4.19)? We finally apply our general theory to the modal approximation. Here the true damping 𝐶 and its modal approximation 𝐶 0 are not equally well known that is, the difference between them will be measured by the modal damping alone, for instance, ∣𝑥∗ (𝐶 − 𝐶 0 )𝑥∣ ≤ 𝜂𝑥∗ 𝐶 0 𝑥,
𝜂 < 1.
(22.10)
As is immediately seen this is equivalent to the same estimate for the damping matrix 𝐷 = 𝛷𝑇 𝐶𝛷 and its modal approximation 𝐷0 = 𝛷𝑇 𝐶 0 𝛷; therefore we have ∥𝐷 ′′ ∥ ≤ 𝜂 with
{ 𝑑′′𝑖𝑗
In other words,
=
0, √ 𝑑𝑖𝑗
𝑑𝑖𝑖 𝑑𝑗𝑗
𝑖 = 𝑗 or 𝑑𝑖𝑖 𝑑𝑗𝑗 = 0 , otherwise.
∣𝑥𝑇 𝐷′′ 𝑦∣ ≤ 𝜂
√
𝑥𝑇 𝐷0 𝑥𝑦 𝑇 𝐷0 𝑦
for any 𝑥, 𝑦. By 𝑥𝑇 𝐷0 𝑥 = 𝑥𝑇 𝐷𝑥 + 𝑥𝑇 (𝐷0 − 𝐷)𝑥 ≤ 𝑥𝑇 𝐷𝑥 + 𝜂𝑥𝑇 𝐷0 𝑥
22 Perturbing Matrix Exponential
191
we obtain 𝑥𝑇 𝐷0 𝑥 ≤
𝑥𝑇 𝐷𝑥 1−𝜂
and finally, for 𝐷′ = 𝐷 − 𝐷 0 ∣𝑥𝑇 𝐷 ′ 𝑦∣ ≤ √
𝜂 √ 𝑇 𝑥 𝐷𝑥𝑦 𝑇 𝐷 0 𝑦. 1−𝜂
Thus, (22.10) implies (22.9) and then (22.6) with 𝜖= √
𝜂 1−𝜂
and the corresponding matrix exponentials are bounded by 𝜂 ∥𝑒𝐵𝑡 − 𝑒𝐴𝑡 ∥ ≤ √ . 2 1−𝜂
Chapter 23
Notes and Remarks
General. The best known monographs on our topic are Lancaster’s book [52] and the work by M¨ uller and Schiehlen [72]; the former is more mathematical than the latter.
Chapter 1 1. Suitable sources on the mechanical background are Goldstein’s book [30] as well as Landau-Livshitz treatise, in particular [61] and [60] where some arguments for the symmetry of the damping matrix are given. A more mathematically oriented (and deeper diving) text is Arnold [3]. 2. Engineering points of view on vibration theory may be found in the monographs [13,42,71]. Rather exhaustive recent works on construction of damping matrices are [76] (with a rich collection of references) and [78].
Chapters 2–4 1. Accurate computation of low (undamped) eigenfrequencies is an important issue, among others because in approximating continuum structures the highest frequencies are present just because one increases the dimension in order to get the lower ones more accurately. This increases the condition number – an old dilemma with any discretisation procedure. We will illustrate this phenomenon on a simple example. Undamped oscillations of a string of length 𝑙 are described by the eigenvalue equation − (𝑎(𝑥)𝑦 ′ )′ = 𝜆𝜌(𝑥)𝑦,
𝑦(0) = 𝑦(𝑙) = 0,
K. Veseli´ c, Damped Oscillations of Linear Systems, Lecture Notes in Mathematics 2023, DOI 10.1007/978-3-642-21335-9 23, © Springer-Verlag Berlin Heidelberg 2011
(23.1)
193
194
23 Notes and Remarks
where 𝑎(𝑥), 𝜌(𝑥) are given positive functions. By introducing the equidistant mesh 𝑙 𝑥𝑗 = 𝑗ℎ, 𝑗 = 1, . . . , 𝑛 + 1, ℎ = 𝑛+1 and the abbreviations 𝑦(𝑥𝑗 ) = 𝑧𝑗 ,
𝑎(𝑥𝑗 ) = 𝑎𝑗 ,
𝜌(𝑥𝑗 ) = 𝜌𝑗 ,
𝑧0 = 𝑧𝑛+1 = 0
we approximate the derivatives by differences (𝑎(𝑥)𝑦 ′ )′𝑥=𝑥𝑗 ≈
𝑎𝑗+1 (𝑧𝑗+1 − 𝑧𝑗 ) − 𝑎𝑗 (𝑧𝑗 − 𝑧𝑗−1 ) ℎ2
(𝑛 + 1)2 (−𝑎𝑗 𝑧𝑗−1 + (𝑎𝑗 + 𝑎𝑗+1 )𝑧𝑗 − 𝑎𝑗+1 𝑧𝑗+1 ). 𝑙2 Thus, (23.1) is approximated by the matrix eigenvalue problem =−
˜ 𝑧, 𝐾𝑧 = 𝜆𝑀
(23.2) 𝜌 𝑙2
𝑗 where 𝐾 is from (1.3) with 𝑘𝑗 = 𝑎𝑗 while 𝑀 is from (1.2) with 𝑚𝑗 = (𝑛+1) 2. For a homogeneous string with 𝑎(𝑥) ≡ 1, 𝜌(𝑥) ≡ 1 both the continuous system (23.1) and its discretisation (23.2) have known eigenvalues
𝜆𝑗 =
𝑗 2 𝜋2 , 𝑙2
2 ˜ 𝑗 = 4 (𝑛 + 1) sin2 𝑗𝜋 . 𝜆 𝑙2 2𝑛 + 2
˜ 𝑗 ≈ 𝜆𝑗 , but the approximation is quite bad for For small 𝑗/𝑛 we have 𝜆 𝑗 close to 𝑛. So, on one hand, 𝑛 has to be large in order to insure good approximation of low eigenvalues but then 𝜅(𝐾) =
˜𝑛 𝜆 ≈ (𝑛 + 1)2 , ˜1 𝜆
(23.3)
which may jeopardise the computational accuracy. With strongly inhomogeneous material (great differences among 𝑘𝑗 or 𝑚𝑗 ) the condition number is likely to be much worse than the one in (23.3). 2. High accuracy may be obtained by (1) using an algorithm which computes the eigenfrequencies with about the same relative error as warranted by the given matrix elements and by (2) choosing the matrix representation on which small relative errors in the matrix elements cause about the same error in the eigenfrequencies. The first issue was treated rather exhaustively in [20] and the literature cited there. The second was addressed in [2], for numerical developments along these lines see, for
23 Notes and Remarks
195
instance [1, 7, 77]. To illustrate the idea of [2] we represent the stiffness matrix 𝐾 of order 𝑛 from Example 1.1 as 𝐾 = 𝐿1 𝐿𝑇1 with ⎡
𝜅1 −𝜅2 0 ⎢ 0 𝜅2 −𝜅3 ⎢ 𝐿1 = ⎢ . ⎣ 0 0 .. 0 0 0
0 ..
.
⎤ 0 0⎥ ⎥ ⎥, 0⎦
𝜅𝑖 =
√
𝑘𝑖 .
𝜅𝑛 𝜅𝑛+1
The matrix 𝐿1 – called the natural factor – is not square but its non-zero elements are close to the physical parameters: square root operation takes place on the stiffnesses (where it is accurate) instead on the stiffness matrix whose condition number is the square of that of 𝐿1 . The eigenfrequencies are the singular values of 𝐿𝑇1 𝑀 −1/2 . From what was said in Sect. 2.2 it is plausible that this way will yield more accurate low frequencies. Not much seems to be known on these issues in the presence of damping. 3. The phase-space construction is even more important for vibrating continua, described by partial differential equations of Mathematical Physics of which our matrix models can be regarded as a finite dimensional approximation. The reader will observe that throughout the text we have avoided the use of the norm of the phase-space matrix since in continuum models this norm is infinite that is, the corresponding linear operator is unbounded. In the continuum case the mere existence and uniqueness of the solution (mostly in terms of semigroups) may be less trivial to establish. Some references for continuous damped systems, their semigroup realisations and properties are [12, 39, 40, 47–49, 73, 90]. The last article contains an infinite dimensional phase space construction which allows for very singular operator coefficients 𝑀, 𝐶, 𝐾. This includes our singular mass matrix case in Chap. 5.In [12] one can find a systematic functionalanalytic treatment with an emphasis on the exponential decay and an ample bibliography. 4. The singular matrix case is an example of a differential algebraic system. We have resolved this system by a careful elimination of ‘nondifferentiated’ variables. Such elimination is generally not easy and a direct treatment is required, see e.g. [10]. 5. Our phase-space matrix 𝐴 is a special linearisation of the second order equation (1.1). There are linearisations beyond the class considered in our Theorem 3.4. In fact, even the phase-space matrices obtained in Chap. 16 are a sort of linearisations not accounted for in Theorem 3.4. A rather general definition of linearisation is the following: We say that a matrix pair 𝑆, 𝑇 or, equivalently, the pencil 𝑆 − 𝜆𝑇 is a linearisation of the pencil 𝑄(𝜆), if [ ] 𝑄(𝜆) 0 = 𝐸(𝜆)(𝑆 − 𝜆𝑇 )𝐹 (𝜆), 0 𝐼
196
23 Notes and Remarks
where 𝐸(𝜆), 𝐹 (𝜆) are some matrix functions of 𝜆 whose determinants are constant in 𝜆. A main feature of any linearisation is that it has the same eigenvalues as the quadratic pencil, including the multiplicities (see [80] and the literature cited there).1 The phase-space matrices, obtained in Chap. 16 do not fall, strictly speaking, under this definition, one has to undo the spectral shift first. The advantage of our particular linearising (1.1) is that it not only enables us to use the structure of dissipativity and 𝐽-symmetry but also gives a proper physical framework of the energy phase space. 6. There is a special linearisation which establishes a connection with models in Relativistic Quantum Mechanics. Equation (1.1) can be written in the form ( ( )2 ) 𝑑 𝐶1 𝐶2 + 𝑦 + 𝐾 − 1 𝑦 = 𝑓1 𝑑𝑡 2 4 with −𝑇 𝐶1 = 𝐿−1 2 𝐶𝐿2 ,
−𝑇 𝐾1 = 𝐿−1 2 𝐾𝐿2 ,
The substitution
( 𝑦1 = 𝑦,
leads to
𝑦2 =
[ ] [ − 𝐶21 𝑑 𝑦1 = 𝑑𝑡 𝑦2 −𝐾1 +
𝐶12 4
𝑓1 = 𝐿−1 2 𝑓.
𝑦 = 𝐿𝑇2 𝑥,
𝑑 𝐶1 + 𝑑𝑡 2
𝐼 − 𝐶21
][
) 𝑦
] [ ] 𝑦1 0 + 𝑦2 𝑓1
Here the phase-space matrix is 𝐽-symmetric with [
] 0𝐼 𝐽= . 𝐼0 This is formally akin to the Klein-Gordon equation which describes spinless relativistic quantum particles and reads (in matrix environment) ((
𝑑 𝑖 +𝑉 𝑑𝑡
)2
) −𝐻
𝑦 = 0,
where 𝑉, 𝐻 are symmetric matrices and 𝐻 is positive definite. The same linearisation leads here to the phase-space equation 𝑖
1 Here,
[ ] [ ] 𝑑 𝑦1 𝑦 =ℋ 1 , 𝑑𝑡 𝑦2 𝑦2
[ ℋ=
𝑉 𝐼 𝐻𝑉
]
too, special care is required, if the mass is singular, see [53].
.
23 Notes and Remarks
197
Note the main difference: here the time is ‘imaginary’ which, of course, changes the dynamics dramatically. Nevertheless, the spectral properties of phase-space matrices are analogous: the overdampedness condition means in the relativistic case the spectral separation (again to the plus and the minus group) which implies the uniform boundedness of the matrix exponential 𝑒−𝑖ℋ𝑡 , −∞ < 𝑡 < ∞, which is here 𝐽-unitary – another difference to our damped system.
Chapters 5–10 1. There are authoritative monographs on the geometry and the spectral theory in the indefinite product spaces. These are Mal’cev [67] and Gohberg, Lancaster and Rodman [27] and more recently [29]. 2. Our Theorem 12.15 has an important converse: if all matrices from a 𝐽-Hermitian neighbourhood of 𝐴 produce only real spectrum near an eigenvalue 𝜆0 of 𝐴 then this eigenvalue is 𝐽-definite. For a thorough discussion of this topic see [27]. 3. Related to Exercise 10.10 is the recent article [35] where the spectral groups of the same sign are connected through the point infinity thus enlarging the class of simply tractable cases. 4. Theorem 10.7 may be used to compute the eigenvalues by bisection which is numerically attractive if the decomposition 𝐽𝐴 − 𝜇𝐽 = 𝐺∗ 𝐽 ′ 𝐺 can be quickly computed for various values of 𝜇. In fact, bisection finds real eigenvalues on a general, not necessarily definitisable 𝐴 whenever the decomposition 𝐽𝐴 − 𝜇𝐽 = 𝐺∗ 𝐽 ′ 𝐺 shows a change of inertia. In order to compute all of them this way one has to know a definitising shift, or more generally, the roots of a definitising polynomial in Theorem 10.3. Indeed, in the latter case the eigenvalues between two adjacent roots are of the same type and by Theorem 10.8 all of them are captured by bisection. 5. Theorem 10.16 is borrowed from the work [46] which contains some further results in this direction.
Chapter 12 1. Jordan canonical forms for 𝐽-Hermitian matrices are given in [57] or [79]. Numerical methods for block-diagonalising are described e.g. in [31]. 2. 𝐽-orthogonal similarities were accepted as natural tool in computing 𝐽-symmetric eigenvalues and eigenvectors long ago, see e.g. [9,11,38,85,88] and in particular [37]. The latter article brings a new short proof of the
198
23 Notes and Remarks
fact that for any 𝐽-orthogonal matrix 𝑈 with a diagonal symmetry 𝐽 the inequality 𝜅(𝐷1 𝑈 𝐷2 ) ≤ 𝜅(𝑈 ) holds for any diagonal matrices 𝐷1 , 𝐷2 (optimal scaling property). This result is akin to our results in Proposition 10.17 and Theorem 12.19. These are taken from [91] where some further results of this type can be found. 3. The computation of an optimal block-diagonaliser, as given in Theorem 12.20 is far from satisfactory because its ‘zeroth’ step, the construction of the initial block-diagonaliser 𝑆, goes via the Schur decomposition which uses a unitary similarity which then destroys the 𝐽-Hermitian property of the initial matrix 𝐴; 𝐽-unitarity of the block diagonaliser is only recovered in the last step. To do the block-diagonalisation in preserving the structure one would want to begin by reducing 𝐴 by a jointly 𝐽, 𝐽 ′ unitary transformation to a ‘triangular-like’ form. Such form does exist (cf. [75]) and it looks like [ [ ] ] 𝑇 𝐵 0𝐼 ′ ′ 𝐴 = , 𝐽 = 0 𝑇∗ 𝐼0 with 𝑇 , say, upper triangular and 𝐵 Hermitian. This form displays all eigenvalues and is attractive to solve the Lyapunov equation, say 𝐴′∗ 𝑋 + 𝑋𝐴′ = −𝐼𝑛 blockwise as 𝑇 ∗ 𝑋11 + 𝑋11 𝑇 = −𝐼𝑛/2 𝑇 ∗ 𝑋12 + 𝑋12 𝑇 ∗ = −𝑋11 − 𝐵 ∗ 𝐵. 𝑇 𝑋22 + 𝑋22 𝑇 ∗ = −𝐼𝑛/2 − 𝐵𝑋12 − 𝑋12
These are triangular Lyapunov/Sylvester equations of half order which are solved in that succession. But there are serious shortcomings: (1) this reduction is possible only in complex arithmetic (except in rare special cases), (2) no real eigenvalues are allowed, unless they have even multiplicities, (3) there are no algorithms for this structure which would beat the efficiency of the standard Schur-form reduction (cf. [22] and the literature cited there). To make a breakthrough here is a hard but important line of research.
Chapter 13 1. Some further informations on the matrix exponential are included in the general monograph [34].
23 Notes and Remarks
199
2. The fact that the knowledge of the spectrum alone is not enough for describing the behaviour of the differential equation was known long ago. A possibility to make the spectrum ‘more robust’ is to introduce so-called ‘pseudospectra’ of different kinds. Rather than a set in ℂ a pseudospectrum is a family of sets. A possible pseudospectrum of a matrix 𝐴 is 𝜎𝜖 (𝐴) = {𝜆 ∈ ℂ : ∥(𝜆𝐼 − 𝐴)−1 ∥ > 1/𝜖}. With 𝜖 → 0 we have 𝜎𝜖 (𝐴) → 𝜎(𝐴) but the speed of this convergence may be different for different matrices and different eigenvalues. So, sometimes for quite small 𝜖 the component of the set 𝜎𝜖 (𝐴) containing some spectral point may still be rather large. The role of the pseudospectra in estimating decay and perturbation bounds is broadly discussed e.g. in [81] and the literature cited there. See also [24, 25].2 3. As it could be expected, our bounds for the exponential decay are mostly much better than those for general matrices (e.g. [44, 64]) because of the special structure of our phase-space matrix. Further improved bounds are given in [74] and [89]. Some comparisons can be found in [89].
Chapter 14 1. The literature on the quadratic eigenvalue problem is sheer enormous. Numerous works are due to Peter Lancaster and his school and co-workers. Few of all these works are cited in our bibliography; a good source for both the information and the biography is the monograph by Lancaster, Gohberg and Rodman [26]. More recent presentations [66] and [80] also include numerical and approximation issues as well as selection of applications which lead to various types of the matrices 𝑀, 𝐶, 𝐾 and accordingly to various structural properties of the derived phase-space matrices. Another general source is [68]. For the properties of 𝐽-definite eigenvalues see also [54]. 2. Numerical methods for our problems have recently been systematically studied under the ‘structural aspect’ that is, one avoids to form any phase-space matrix and sticks at the original quadratic pencil, or else identifies the types of the symmetry of the phase-space matrix and develops ‘structured’ numerical methods and corresponding sensitivity theory. Obviously, ‘structure’ can be understood at various levels, from coarser to finer ones. This research has been done intensively by the school of Volker Mehrmann, substantial information, including bibliography, can
2 The
Russian school uses the term spectral portrait instead of pseudospectrum.
200
23 Notes and Remarks
be found e.g. in [22, 41, 65]. Several perturbation results in the present volume are of this type, too. 3. For quadratic/polynomial numerical ranges see e.g. [56, 69]. A related notion is the ‘block-numerical range’, see e.g. [62, 63, 82]. All these numerical ranges contain the spectrum just like Gershgorin circles or our stretched Cassini ovals in Chap. 19. 4. Concerning overdamped systems see the pioneering work by Duffin [21] as well as [52]. Recent results concerning criteria for overdampedness can be found in [33, 36]. 5. (Exercise 14.10) More on linearly independent eigenvectors can be found in [26].
Chapter 20 A rigorous proof of Theorem 20.9 is found in [27].
Chapter 21 1. For results on the sensitivity (condition number) of the Lyapunov operator (21.6) see [6] and references there. A classical method for solving Lyapunov and Sylvester equations with general coefficient matrices is given in [5]. For more recent results see [43, 50]. 2. Theorem 21.10 is taken from [16] where also some interesting phenomena about the positioning of dampers are observed. The problem of optimal positioning of dampers seems to be quite hard and it deserves more attention. Further results of optimal damping are found in [14,15]. Related to this is the so called ‘eigenvalue assignment’ problem, see [18] and the literature cited there. The statement of Theorem 21.10 can be generalised to the penalty function 𝑋 → Tr(𝑋𝑆), where 𝑆 is positive semidefinite and commutes with 𝛺, see the dissertation [73] which also contains an infinite dimensional version and applications to continuous damped systems. 3. For references on inverse spectral theory, applied to damped systems see e.g. the works of Lancaster and coauthors [23,55,59]. A result on an inverse spectral problem for the oscillator ladder with just one damper in Example 1.1 is given in [86, 87]. In this special case it was experimentally observed, (but yet not proved) that the spectral abscissa and the minimal trace
23 Notes and Remarks
201
criterion (21.4) yield the same minimiser (independently of the number of mass points). 4. An efficient algorithm for solving the Lyapunov equation with rank-one damping is given in [87]. Efficient solving and trace-optimising of the Lyapunov equation is still under development, see e.g. [83, 84] and the literature cited there.
Chapter 22 Our perturbation bounds are especially designed for dissipative or a least asymptotically stable matrices. As could be expected most common bounds are not uniform in 𝑡 (see e.g. [64]).
References
1. Amestoy, P. R., Duff, I. S., and Puglisi, C., Multifrontal QR factorization in a multiprocessor environment. Technical Report TR/PN94/09, ENSEEIHT, Toulouse, France 1994 2. Argyris, J. H., Brønlund, O. E., The natural factor formulation of stiffness for the matrix deplacement method, Comp. Methods Appl. Mech. Engrg. 40 (1975) 97–119 3. Arnol´d, V. I., Mathematical methods of classical mechanics, Springer, New York (1989) 4. Ashcraft, C., Grimes, R., Lewis, J., Accurate Symmetric Indefinite Linear Equation Solvers, SIMAX 20 (1999), 513–561 5. Bartels, R. H., Stewart, G. W., A solution of the matrix equation AX + XB = C, Comm. ACM 15 (1972) 820–826 6. Bhatia, R., A note on the Lyapunov equation, LAA 259 (1997) 71–76 7. Boman, E. G., Hendrickson, B., Vavasis, S., Solving elliptic finite element systems in near-linear time with support preconditioners, arXiv:cs/0407022v4 [cs.NA] 4 Apr 2008 8. Brauer, A. Limits for the characteristic roots of a matrix, Duke Math. J. 13 (1946) 387–395 9. Brebner, M. A., Grad, J., Eigenvalues of Ax = λBx for real symmetric matrices A and B computed by reduction to a pseudosymmetric form on the HR process, LAA 43 (1982) 99–118 10. Brenan, K. E., Campbell, S. L., and Petzold, L. R., Numerical Solution of Initial-Value Problems in Differential Algebraic Equations. Elsevier, North Holland (1989) 11. Bunse-Gerstner, A., An analysis of the HR algorithm for computing the eigenvalues of a matrix, LAA 35 (1981) 155–173 12. Chen, G., Zhou, J., Vibration and Damping in Distributed Systems I, II CRC Press (1993) 13. Clough, R. W., Penzien, J., Dynamics of structures, McGraw-Hill, NY (1993) 14. Cox, S. J., Designing for optimal energy absorption, III: Numerical minimization of the spectral abscissa, Structural and Multidisciplinary Optimization, 13 (1997) 17–22 15. Cox, S. J., Designing of optimal energy absorption, I: Lumped parameter systems, ASMA J. Vib. and Acoustics, 120 (1998) 339–345 16. Cox, S. J., Naki´c, I., Rittmann, A., Veseli´c, K., Lyapunov optimization of a damped system, Systems and Control Letters, 53 (2004) 187–194 17. Danisch, R., Delini´c, K., Veseli´c, K., Quadratic eigenvalue solver for modal response analysis of non-proportionally damped elastic structures, Trans. 7-th Int. Conf. on Struct. Mech. in Reactor Technology, Chicago 1983, K 3/1, North-Holland (1983)
K. Veseli´c, Damped Oscillations of Linear Systems, Lecture Notes in Mathematics 2023, DOI 10.1007/978-3-642-21335-9, © Springer-Verlag Berlin Heidelberg 2011
203
204
References
18. Datta, B. N., Numerical methods for linear control systems. Design and analysis. Elsevier (2004) 19. Demmel, J., Veseli´ c, K., Jacobi’s method is more accurate than QR, SIAM J. Matrix Anal. Appl., 13 (1992), 1204–1245 20. Drmaˇc, Z., Veseli´c, K., New fast and accurate Jacobi algorithm I, SIMAX 299 (2008) 1322–1342; II, SIMAX 30 (2008) 1343–1362 21. Duffin, R. J., A minimax theory for overdamped networks, J. Rational Mech. Anal. 4 (1955), 221–233 22. Fassbender, H., Kressner, D., Structured eigenvalue problems, GAMM-Mitt. 29 (2006) no. 2, 297–318 23. Freitas, P., Lancaster, P., Optimal decay of energy for a system of linear oscillators, SIAM J. Matrix Anal. Appl. 21 (1999), no. 1, 195–208 24. Godunov, S. V., Lectures on modern aspects of Linear Algebra (Russian), N. Kniga (2002) 25. Godunov, S. V., Kiriljuk, O. P., Kostin, I. V., Spectral portraits of matrices (Russian), AN SSSR Siber. Otd. Novosibirsk (1990) (preprint) 26. Gohberg, I., Lancaster, P., Rodman, L., Matrix polynomials, Academic, New York (1982) 27. Gohberg, I., Lancaster, P., Rodman, L., Matrices and indefinite scalar products, Birkh¨ auser, Basel (1983) 28. Gohberg, I., Lancaster, P., Rodman, L., Quadratic matrix polynomials with a parameter. Adv. in Appl. Math. 7 (1986), 253–281 29. Gohberg, I., Lancaster, P., Rodman, L., Indefinite linear algebra and applications, Springer (2005) 30. Goldstein, H., Classical Mechanics, Addison-Wesley (1959) 31. Golub, G. H., van Loan, Ch. F., Matrix computations, J. Hopkins University Press, Baltimore (1989) 32. Guang-Da Hu and Guang-Di Hu, A relation between the weighted logarithmic norm of a matrix and the Lyapunov equation, BIT 40 (2000) 606–610 33. Guo, Chun-Hua, Lancaster, P., Algorithms for hyperbolic quadratic eigenvalue problems, Math. Comp. 74 (2005) 1777–1791 34. Higham, N. J., Functions of Matrices: Theory and Computation. SIAM (2008) 35. Higham, N. J., Mackey, S. D., Tisseur, F., Definite matrix polynomials and their linearization by definite pencils, University of Manchester MIMS EPrint 2007.97 36. Higham, N. J., van Dooren, P., Tisseur, F., Detecting a definite Hermitian pair and a hyperbolic or elliptic quadratic eigenvalue problem and associated nearness problems, Linear Algebra Appl. 351–352 (2002) 455–474 37. Higham, N. J., J-orthogonal matrices: Properties and generation, SIAM R. (2003) 504–519 38. Hoppe, P., Jacobi-¨ ahnliche Blockverfahren zur numerischen Behandlung des J-symmetrischen Eigenwertproblems, PhD Fernuniversit¨ at Hagen 1984. http://www.fernuni-hagen.de/MATHPHYS/veselic/doktorandi.html 39. Hryniv, R., Shkalikov, A., Operator models in elasticity theory and hydrodynamics and associated analytic semigroups, Vestnik Moskov. Univ. Ser. I Mat. Mekh. (1999), no. 5, 5–14 40. Hryniv, R., Shkalikov, A., Exponential stability of semigroup s associated with some operator models in mechanics. (Russian) Mat. Zametki 73 (2003), no. 5, 657–664 41. Hwang, T-M., Lin, W-W., Mehrmann, V., Numerical solution of quadratic eigenvalue problems with structure-preserving methods, SIAM J. Sci. Comput. 24 (2003) 1283– 1302 42. Inman, D. J., Vibration with Control Measurement and Stability, Prentice Hall (1989) 43. Jonsson, I., K˚ agstr¨ om, B., Recursive blocked algorithm for solving triangular systems. I. One-sided and coupled Sylvester-type matrix equations, ACM Trans. Math. Software, 28 (2002) 392–415
References
205
44. K˚ agstrom, B., Bounds and Perturbation Bounds for the Matrix Exponential, BIT 17 (1977) 39–57 45. Knowles, J. K., On the approximation of damped linear systems, Struct. Control Health Monit. 13 (1995) 324–335 46. Kovaˇ c-Striko, M., Veseli´c, K., Trace minimization and definiteness of symmetric matrix pairs, LAA 216 (1995) 139–158 47. Kre˘ın, M. G., Langer, G. K., Certain mathematical principles of the linear theory of damped vibrations of continua. (Russian) 1965 Appl. Theory of Functions in Continuum Mechanics (Proc. Internat. Sympos., Tbilisi, 1963), Vol. II, Fluid and Gas Mechanics, Math. Methods (Russian) pp. 283–322 Izdat. “Nauka”, Moscow 48. Kre˘ın, M. G., Langer, H., On some mathematical principles in the linear theory of damped oscillations of continua. I. Translated from the Russian by R. Troelstra, Integral Equations Operator Theory 1 (1978), no. 4, 539–566 49. Kre˘ın, M. G., Langer, H., On some mathematical principles in the linear theory of damped oscillations of continua. II. Translated from the Russian by R. Troelstra, Integral Equations Operator Theory 1 (1978), no. 3, 364–399 50. Kressner, D., Block variants of Hammarling’s method for solving Lyapunov equations, ACM Trans. Math. Software 34 (2008) 1–15 51. Lancaster, P., Quadratic eigenvalue problems, LAA 150 (1991) 499–506 52. Lancaster, P., Lambda-matrices and vibrating systems, Pergamon Press Oxford 1966, Dover (2002) 53. Lancaster, P., Linearization of regular matrix polynomials, Electronic J. Lin. Alg. 17 (2008) 21–27 54. Lancaster, P., Model updating for self-adjoint quadratic eigenvalue problems, LAA 428 (2008) 55. Lancaster, P., Prells, U., Inverse problems for damped vibrating systems, J. Sound Vibration 283 (2005) 891–914 56. Lancaster, P., Psarrakos, P., The Numerical Range of Self-Adjoint Quadratic Matrix Polynomials, SIMAX bf 23 (2001) 615–631 57. Lancaster, P., Rodman, L., Canonical forms for hermitian matrix pairs under strict equivalence and congruence, SIAM Review, bf 47 (2005) 407–443 58. Lancaster, P., Tismenetsky, M., The theory of matrices, Academic (1985) 59. Lancaster, P., Ye, Q., Inverse spectral problems for linear and quadratic matrix pencils. Linear Algebra Appl. 107 (1988), 293–309 60. Landau, L. D., Lifshitz, E. M., Course of theoretical physics. Vol. 5: Statistical physics. Pergamon Press, Oxford (1968) 61. Landau, L. D., Lifshitz, E. M., Course of theoretical physics. Vol. 1. Mechanics. Pergamon Press, Oxford (1976) 62. Langer, H., Matsaev, V., Markus, A., Tretter, C., A new concept for block operator matrices: The quadratic numerical range, Linear Algebra Appl. 330 (2001) 89–112 63. Langer, H., Tretter, C., Spectral decomposition of some block operator matrices, J. Operator Theory 39 (1998) 1–20 64. van Loan, Ch., The Sensitivity of the Matrix Exponential, SIAM J. Num. Anal. 13 (1977) 971–981 65. Mackey, D. S., Mackey, N., Mehl, C., Mehrmann, V., Structured Polynomial Eigenvalue Problems: Good Vibrations from Good Linearizations, SIMAX 28 (2006) 1029–1051 66. Mackey, D. S., Mackey, N., Mehl, C., Mehrmann, V., Vector Spaces of Linearizations for Matrix Polynomials, SIMAX 28 (2006) 971–1004 67. Mal’cev, A. I., Foundations of linear algebra. Translated from the Russian by Thomas Craig Brown; edited by J. B. Roberts W. H. Freeman & Co., San Francisco, xi+304 (1963) 68. Markus, A. S., Introduction to the spectral theory of polynomial operator pencils, AMS (1988)
206
References
69. Markus, A. S., Rodman, L. Some results on numerical ranges and factorizations of matrix polynomials, Linear and Multilinear Algebra 42 (1997), 169–185 70. Mehrmann, V., The autonomous linear quadratic control problem: theory and numerical solution, Lecture notes in control and information sciences 163 Springer (1977) 71. Meirovich, L., Elements of vibration analysis, McGraw-Hill, NY (1984) 72. M¨ uller, P. C., Schiehlen, W. O., Lineare Schwingungen, Akademische Verlagsgesellschaft Wiesbaden (1976) 73. Naki´c, I., Optimal damping of vibrational systems, Dissertation, Fernuniversit¨ at Hagen (2003). www.fernuni-hagen.de/MATHPHYS/veselic/doktorandi.html 74. Nechepurenko, Yu. M., On norm bounds for the matrix exponential, Doklady Mathematics, 63 (2001) 233–236 75. Paige, C., van Loan, C., A Schur decomposition for Hamiltonian matrices, LAA, 41 (1981), 11–32 76. Pilkey, D., Computation of a damping matrix for finite element model updating, Ph. D. dissertation (1998). citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.2.5658 77. Plemmons. R. J., Minimum norm solutions to linear elastic analysis problems, Int. J. Num. Meth. Engrg. 20 (2005) 983–998 78. Srikantha Phania, A., Woodhouse, J., Viscous damping identification in linear vibration, Journal of Sound and Vibration 303 (2007), 475–500 79. Thompson, R. C., Pencils of complex and real symmetric and skew matrices, Linear Algebra Appl. 147 (1991) 323–371 80. Tisseur, F., Meerbergen, K., The Quadratic Eigenvalue Problem, SIAM R. 43 235–286 (2001) 81. Trefethen, L. N., Embree, M., Spectra and pseudospectra, Princeton University Press (2005) 82. Tretter, Ch., Spectral Theory of Block Operator Matrices and Applications, Imperial College Pr. (2008) 83. Truhar, N., An efficient algorithm for damper optimization for linear vibrating systems using Lyapunov equation, J. Comp. Appl. Math. 172 (2004) 169–182 84. Truhar, N., Veseli´c, K., An efficient method for estimating the optimal dampers’ viscosity for linear vibrating systems using Lyapunov equation, 2007, SIMAX (to appear) 85. Veseli´c, K., Global Jacobi method for a symmetric indefinite problem Sx = λT x, Comp. Meth. Appl. Mech. Engrg. 38 (1983) 273–290 86. Veseli´c, K., On linear vibrational systems with one dimensional damping, Appl. Anal. 29 (1988) 1–18 87. Veseli´c, K., On linear vibrational systems with one dimensional damping II, Integral Eq. Operator Th. 13 (1990) 883–897 88. Veseli´c, K., A Jacobi eigenreduction algorithm for definite matrix pairs, Numer. Math. 64 (1993) 241–269 89. Veseli´c, K., Bounds for exponentially stable semigroups, Linear Algebra Appl. 358 (2003) 309–333 90. Veseli´c, K., Bounds for contractive semigroups and second order systems, Operator Theory: Advances and Applications, 162 (2005) 293–308 91. Veseli´c, K:, Slapniˇcar, I., On spectral condition of J-Hermitian operators, Glasnik Mat. 35(55) (2000) 3–23
Index
adjoint eigenvalue problem, 132 approximation maximal modal, 152 modal, 152 proportional, 148 average amplitude displacement, 141 energy, 141
block-diagonaliser, 106
Cassini ovals double, 154 modified, 157 stretched, 147 complement J-orthogonal, 46 configuration, 6 space, 6 conservative system, 6 coordinate transformation, 20
damped modally, 22 system, 1 damper positioning, 200 decomposition complex spectral, 102 indefinite, 42 J,J’-unitary, 100 J-orthogonal, 63 J-orthogonal, of the identity, 63 spectral, 97
definiteness interval, 76 definitising shift, 76 departure from normality, 27 differential algebraic system, 195
eigenfrequencies (undamped), 16 eigenvalue decomposition generalised, 15 standard, 15 eigenvalues, xv (non-)defective, 99 J-definite, 73 J-negative, 73 J-positive, 73 of mixed type, 73 of positive/negative/definite type, 73 semisimple, 99 energy balance, 6 dissipation, 6 kinetic, 5 potential, 5 total, 6 energy phase space, 25 equation evolution, 25 Lyapunov, 117 Lyapunov, standard, 178 Sylvester, 118 equilibrium, 7
field of values, 129 force generalised external, 7 harmonic, 139
K. Veseli´c, Damped Oscillations of Linear Systems, Lecture Notes in Mathematics 2023, DOI 10.1007/978-3-642-21335-9, © Springer-Verlag Berlin Heidelberg 2011
207
208
Index
inertial, 4 formula Duhamel, 24 minimax, Duffin’s, 163 minimax, generalised, 17 minimax, Hermitian, 84 minimax, J-Hermitian, 82
phase-space, 25 quadratic pencil, 122 eigenspace, 122 overdamped, 125 resistance, 12 surjective, xiii modal representation, 26
inertia, 45 inverse spectral problem, 181
natural factor, 195 non-degenerate scalar product, 40 numerical range, 129 polynomial, 130 quadratic, 130
J-orthogonal companion, 46 sets, 41 vectors, 40
Lagrange equations, 7 Lagrangian, 7 linearisation of a non-linear system, 8 of a second order system, 26 Lyapunov equation, 117 Lyapunov operator, 170 Lyapunov solution, 119
matrix (J-)definitisable, 76 asymptotically stable, 117 dissipative, 25 exponential, contractive, 25 exponential, exponentially decaying, 117 inductance, 12 injective, xiii inverse capacitance, 13 J,J’-isometry, 52 J,J’-orthogonal, 52 J,J’-unitarily diagonalisable, 69 J,J’-unitary, 52 J-adjoint, 49 J-Hermitian, 49 J-orthogonal, 49 J-symmetric, 25, 49 J-transpose, 49 J-unitary, 49 jointly J-Hermitian, 54 jointly unitary, 50 nilpotent, 99 pair, definitisable, 76 pencil, 15
oscillations steady state, 141 transient, 141
perfect shuffling, 104 polynomial definitising, 75 projection, 55 J-orthogonal, 61 J-positive/negative, 65 spectral, 96 pseudospectra, 199
Rayleigh quotient, 79 relative stiffness, 17 resolvent, 94 of a quadratic pencil, 140 resonance, 140 response harmonic, 139 of a linear system, 1 root space, 97
sign group, 74 sign partition, 75 spectral abscissa, 116 spectral points, xv spectral portrait, 199 spectrum definite, 74 spring stiffness, 3 symmetry, 40
trajectory, 187
Index undamped frequencies, 16 system, 15
vector J-definite, 41 J-negative, 41
209 J-neutral, 41 J-normalised, 41 J-positive, 41 viscosities, 3 viscosity relation, 126
work, 4
LECTURE NOTES IN MATHEMATICS
123
Edited by J.-M. Morel, B. Teissier; P.K. Maini Editorial Policy (for the publication of monographs) 1. Lecture Notes aim to report new developments in all areas of mathematics and their applications - quickly, informally and at a high level. Mathematical texts analysing new developments in modelling and numerical simulation are welcome. Monograph manuscripts should be reasonably self-contained and rounded off. Thus they may, and often will, present not only results of the author but also related work by other people. They may be based on specialised lecture courses. Furthermore, the manuscripts should provide sufficient motivation, examples and applications. This clearly distinguishes Lecture Notes from journal articles or technical reports which normally are very concise. Articles intended for a journal but too long to be accepted by most journals, usually do not have this “lecture notes” character. For similar reasons it is unusual for doctoral theses to be accepted for the Lecture Notes series, though habilitation theses may be appropriate. 2. Manuscripts should be submitted either online at www.editorialmanager.com/lnm to Springer’s mathematics editorial in Heidelberg, or to one of the series editors. In general, manuscripts will be sent out to 2 external referees for evaluation. If a decision cannot yet be reached on the basis of the first 2 reports, further referees may be contacted: The author will be informed of this. A final decision to publish can be made only on the basis of the complete manuscript, however a refereeing process leading to a preliminary decision can be based on a pre-final or incomplete manuscript. The strict minimum amount of material that will be considered should include a detailed outline describing the planned contents of each chapter, a bibliography and several sample chapters. Authors should be aware that incomplete or insufficiently close to final manuscripts almost always result in longer refereeing times and nevertheless unclear referees’ recommendations, making further refereeing of a final draft necessary. Authors should also be aware that parallel submission of their manuscript to another publisher while under consideration for LNM will in general lead to immediate rejection. 3. Manuscripts should in general be submitted in English. Final manuscripts should contain at least 100 pages of mathematical text and should always include – a table of contents; – an informative introduction, with adequate motivation and perhaps some historical remarks: it should be accessible to a reader not intimately familiar with the topic treated; – a subject index: as a rule this is genuinely helpful for the reader. For evaluation purposes, manuscripts may be submitted in print or electronic form (print form is still preferred by most referees), in the latter case preferably as pdf- or zipped psfiles. Lecture Notes volumes are, as a rule, printed digitally from the authors’ files. To ensure best results, authors are asked to use the LaTeX2e style files available from Springer’s web-server at: ftp://ftp.springer.de/pub/tex/latex/svmonot1/ (for monographs) and ftp://ftp.springer.de/pub/tex/latex/svmultt1/ (for summer schools/tutorials).
Additional technical instructions, if necessary, are available on request from
[email protected]. 4. Careful preparation of the manuscripts will help keep production time short besides ensuring satisfactory appearance of the finished book in print and online. After acceptance of the manuscript authors will be asked to prepare the final LaTeX source files and also the corresponding dvi-, pdf- or zipped ps-file. The LaTeX source files are essential for producing the full-text online version of the book (see http://www.springerlink. com/openurl.asp?genre=journal&issn=0075-8434 for the existing online volumes of LNM). The actual production of a Lecture Notes volume takes approximately 12 weeks. 5. Authors receive a total of 50 free copies of their volume, but no royalties. They are entitled to a discount of 33.3 % on the price of Springer books purchased for their personal use, if ordering directly from Springer. 6. Commitment to publish is made by letter of intent rather than by signing a formal contract. Springer-Verlag secures the copyright for each volume. Authors are free to reuse material contained in their LNM volumes in later publications: a brief written (or e-mail) request for formal permission is sufficient. Addresses: Professor J.-M. Morel, CMLA, ´ Ecole Normale Sup´erieure de Cachan, 61 Avenue du Pr´esident Wilson, 94235 Cachan Cedex, France E-mail:
[email protected] Professor B. Teissier, Institut Math´ematique de Jussieu, ´ UMR 7586 du CNRS, Equipe “G´eom´etrie et Dynamique”, 175 rue du Chevaleret 75013 Paris, France E-mail:
[email protected] For the “Mathematical Biosciences Subseries” of LNM: Professor P. K. Maini, Center for Mathematical Biology, Mathematical Institute, 24-29 St Giles, Oxford OX1 3LP, UK E-mail :
[email protected] Springer, Mathematics Editorial, Tiergartenstr. 17, 69121 Heidelberg, Germany, Tel.: +49 (6221) 487-8259 Fax: +49 (6221) 4876-8259 E-mail:
[email protected]